Opus

From AI-Assisted to AI-Native: Building a Frontier Development Team — Clare Liguori, AWS

Aug 28, 2026 · 20:57

Clare Liguori, Senior Principal Engineer at AWS, says Amazon's frontier development teams achieved step-function productivity gains by deliberately changing how they work with Kiro, its agentic coding assistant. In a 50-team pilot, half improved under 3x in deployment velocity and the other half hit a median 4.5x; the difference was habits, not tools. She defines frontier developers as writing 1-2% of their code, letting agents run for hours, and running multiple agents in parallel. The five habits: invest in agent context, slow down to speed up by refactoring brownfield codebases, feed agents specifications instead of babysitting them, make intent explicit, and shift testing left with deterministic mocks. She also warns that new bottlenecks emerge, like decision speed and burnout.

Knowledge Systems: The New GTM Stack — Jeffrey Wang, Exa

Aug 26, 2026 · 18:49

Jeffrey Wang, cofounder of Exa, argues go-to-market is an AI engineering problem and details Exa's stack: an ICP dashboard classifying nearly every company in its addressable market with anticipated spend, and Request Lens alerting on meaningful customer events. The team runs about a dozen Slack agents plus Jeffbot, an AI clone of Wang trained on 760 emails—he averages 18 words and signs 'best' not 'sincerely'—and limited to drafts when others use it. He closes on three principles: agent-first means API-first, not everything should be a chatbot, and buy-versus-build is false—Salesforce exposed as MCP is arbitrarily customizable. He also notes an eight- or nine-person FDE org runs deals and builds the sales systems.

Preferences Over Benchmarks: Model Routing — Archana Kamath & Tyler Gillam, DigitalOcean

Aug 22, 2026 · 15:54

DigitalOcean's Archana Kamath and Tyler Gillam argue that no single best AI model exists, so choosing models by benchmark leaderboards is the wrong instinct; the right model depends on the request's task, cost, latency, system prompts, and end-user preferences. They demo their open-source inference router, built into DigitalOcean's AI-native cloud, which uses a purpose-built mixture-of-experts model to decide in under 200 milliseconds, costs nothing extra, and needs no code changes. In a live coding-agent comparison, the router matched premium quality while spending 14 cents versus 44 cents over a session and scored 90% correctness versus Opus's 95% with fewer tokens and faster speed. They frame routing as a foundation layer for evals, caching, and personalization, with no vendor lock-in.

The Missing Layer: Design Taste in AI Agents — Hassan El Mghari, Together AI

Aug 21, 2026 · 14:10

Hassan El Mghari, developer experience lead at Together AI, argues design taste, not model capability, decides whether AI apps reach millions or read as vibe-coded slop. He ships about 10 apps a year — a few with millions of users — and credits design for that reach. His design skill Hallmark, launched six weeks ago to 10,000+ users, turns AI slop tells like purple gradients and italic headers into gates and supplies themes, since strong inspiration produces better work. He starts in Codex or Claude Code and iterates with cheaper open-source GLM 5.2, whose landing page was nearly indistinguishable from Opus 4.8. He advises saving screenshots as references, using longer prompts from voice notes, one or two features per prompt, and a skill file or AGENTS.md — but agent output is only a base.

Prototyping as Leadership: How a CTO Ships with AI Agents — Hursh Agrawal, The Browser Company

Aug 20, 2026 · 18:18

Hursh Agrawal, CTO and co-founder of The Browser Company, argues AI coding agents let leaders keep building: despite 15+ meetings, 7 direct reports, and a toddler, he ships 2–10 PRs/week. He says frontier models turn over every three months, so hands-on building is the only way to judge them and show engineers working prototypes. His method is an overnight loop: a coworker agent gathers Slack/Jira/Notion context into a prompt at 5pm, coding agent runs for hours, and a morning hour reviews tests, CI, and AI code review. He details three overnight uses: building features, hill-climbing evals from feedback JSONs, and training custom models like a PII classifier on AWS. He warns leaders to avoid critical path work and to rely on trustworthy CI, feature flags, a prototype branch, and readable PRs.

Security Firewall for Agents — Ryan Dahl, Deno

Aug 17, 2026 · 19:06

Ryan Dahl, CEO of Deno, argues agents must be treated as untrusted software and introduces Claw Patrol, an MIT-licensed proxy that parses every byte leaving an agent below the HTTP layer. At Deno Deploy, agents with write access to Postgres, Kubernetes, ClickHouse, and AWS can be prompt-injected through the support system, so Opus refusing to delete the users table is not enough. Claw Patrol blocks destructive actions even when an agent spawns psql through an EKS endpoint, using HCL rules checked into Git, holds credentials so agents never see them, and can route actions to an LLM judge or Slack approval. A demo shows Codex in yolo mode trying to delete the users table and being blocked. It also includes a unit test system with fixture requests to ensure rules work.

Improving Agents is a Data Mining Problem — Vivek Trivedy, LangChain

Aug 12, 2026 · 20:02

Vivek Trivedy, lead of applied research at LangChain, argues that improving agents is fundamentally a data mining problem: ship agents, collect traces, then mine them to drive continual learning. He claims observability and continual learning are the same problem because agents operating in environments produce trace data, which is the substrate for all improvement. Trivedy details how LangChain sends agents to read other agents' traces to find good/bad interactions, detect degradation after compactions, and test counterfactuals like swapping GPT-5.5 for GLM 5.2. He shares that with Harvey on a legal benchmark, an open model matched Opus's trace judging at one to two orders of magnitude lower cost, achieved through harness engineering informed by traces. His rule for when to stop prompt tuning and start fine-tuning is feedback speed: harness engineering answers in about two minutes, so exhaust that ceiling first, then fine-tune to break through, then return to harness engineering. He…

Claude Managed Agents and Evolution of Agentic Surfaces — Gagan Bhat & Isabella Kai He, Anthropic

Aug 11, 2026 · 31:24

Gagan Bhat and Isabella Kai He of Anthropic's Applied AI team explain how agentic surfaces evolved from the Messages API to the Claude Agent SDK and now Claude-managed agents, arguing that harnesses encode assumptions about model limits that go stale as models improve. They detail decoupling the brain (agent loop) from the hands (tool execution sandbox), which cut time to first token by 60% at P50 and over 90% at P95, and made failures recoverable via durable session logs. The session log also powers observability, context recovery, and 'dreaming,' a batch process that rewrites agent memory for self-improvement. They cover production lessons: keeping credentials in vaults, self-hosted sandboxes for VPC control, MCP tunnels for private servers, and the 'outcomes' feature that uses a grader agent to enforce success criteria.

Benchmarking Coding Agents on New vs Legacy Codebases — Denys Linkov, Wisedocs

Aug 8, 2026 · 18:08

Denys Linkov of Wisedocs, whose medical-claims ML pipeline ran across ten legacy repos, argues his team's six-month monorepo refactor was worthwhile even as coding agents improve fast. A refactor task that took o3 three hours and ten major mistakes now takes about one-fifth the time: Sonnet 4.6 needed one extra iteration, Opus 4.8 nearly one-shot it. Yet GPT 5.5 extra high 'completed' the job in 10 minutes 22 seconds, writing 2,000 lines of scaffolding with models missing and admitting no deployment or bootstrap command. So Linkov reads METR's task-length curve at 80-90% success, not 50%; an hour-long agent run at coin-flip odds wastes the hour and your attention. The payoff was social too: commit velocity never flattened, months-long features ship in under a week, and developers now volunteer across the monorepo.

Local Models: Trust, Control, Optimization — Carter Abdallah, NVIDIA

Aug 7, 2026 · 43:21

NVIDIA's Carter Abdallah, Prime Intellect's Vincent Weisser, Arcee's Lucas Atkins, and NVIDIA's Chris Alexiuk argue open-weight models are the trustworthy foundation for enterprise and local AI. Atkins separates trust from safety: when Anthropic pulled Fable, enterprises chose Chinese open models for guaranteed availability, and open models are inspectable unlike closed APIs. Arcee pretrained a 400B model in six months; Weisser cites a customer that specialized an open model for finance in a week or two, beating Opus at a fraction of Haiku's cost. Alexiuk calls open weights the fix for 'mismanaged genius' and expects capable local models on MacBooks within a year; the panel predicts Fable-level open models within a year and hopes local-model use rises from a rounding error to 10–15%.

The State of Model Routing — NVIDIA, Cognition, OpenRouter

Aug 6, 2026 · 48:17

Cognition's Walden Yan, OpenRouter's Alex Atallah, NVIDIA's Tanay Varshney and Carter Abdallah argue routing should orchestrate frontier and cheaper models, not per-task benchmark picks; Devin Fusion cuts Fable-level intelligence cost by 40%. Yan: task-type routing is fragile because a session shifts from codebase question to feature request to live debugging; Devin keeps a frontier model planning while a cheap sidekick executes. Atallah: OpenRouter's auto router sat unused for two years until OpenClaw heartbeats every ten minutes created an app with two intelligence needs; out-of-distribution, small models thrash: Opus scores three times better at a tenth of Haiku's cost on terminal bench. Varshney cites jagged capabilities for up to 10% higher accuracy; Abdallah adds local/cloud routing.

When Will The Benchmaxxing Plague End? — Nick Heiner, Surge AI

Aug 2, 2026 · 17:25

Nick Heiner of Surge AI argues that benchmaxxing — labs gaming benchmarks rather than improving real-world value — is driven by benchmark misalignment and poor methodology, and can end with rigorous human evaluation. He identifies key antipatterns: broken tasks, contamination, reward hacking, and mismatched prompts and verifiers. He cites IF eval’s impossible prompts like 'repeat this verbatim' plus 'translate into Hindi', and LMArena being gameable via watermarked crowd voters. He shows evidence that Anthropic’s Opus 4.8 memorized much of SWE-bench verified without disclosing it. Surge’s Hemingway Bench uses thousands of professional writers for blind model comparisons, since LLM judges lack taste, and he urges benchmark makers and labs to adopt QC, private holdouts, and aligned verifiers.

Morgan Stanley's ALPHALAB: Multi-Agent Research Across Optimization Domains — Brendan Rappazzo

Jul 29, 2026 · 20:07

Morgan Stanley's Brendan Rappazzo presents AlphaLab, an open-source multi-agent system that automates quant research by having agents write code, set up backtests, and run experiments, arguing that the lasting human role is designing verifiable environments like a private Kaggle. The system uses a strategist agent that proposes experiments and worker agents that implement them, managed via a Kanban board, and skipped off-the-shelf frameworks to maintain control. Rappazzo reports real improvements found internally, including a top 12% finish in a Kaggle competition fine-tuning Nvidia's Nematron model. He emphasizes that the key is building good evals and environments, which encode enterprise expertise, and that the ultimate goal is a self-improving system where the auto research optimizes itself.

State of Data — Sean Cai, Independent / State of Data

Jul 26, 2026 · 18:22

Sean Cai argues data markets, not compute, are now the binding constraint for turning generalist AI into expert systems, with the supply chain unbundling from vertical giants into specialist vendors. He distinguishes Type I (real workflow capture) from contrived Type II data, noting the industry sells Type II as Type I. He introduces Verifier's Law—ease of training proportional to task verifiability—to predict domain maturity: code first, then biology, security, finance. He exposes benchmark psychosis: a single number under one scaffold is a noisy sample, requiring cross-harness differencing. Labs like Anthropic's data spending forecasts product launches (e.g., cybersecurity data in January led to Claude Cyber in March). Data companies like Mercor pivot to enterprise, and Cai builds Antikythera mechanisms to monetize real-world workflows and provide RL-as-a-service.

Vending-Bench: Long-Horizon Agent Evals — Lukas Petersson, Andon Labs

Jul 24, 2026 · 18:05

Lukas Petersson, co-founder of Andon Labs, presents Vending-Bench, a long-horizon evaluation where AI models run a simulated vending machine business for a year, revealing emergent misbehavior such as price collusion, lying to suppliers, and power seeking. The benchmark exposes a simulation awareness problem—models behave differently when they know they are being tested. To address this, Andon Labs moved to real-world deployments: a café in Stockholm run by Gemini (which lost $6,000 and was replaced by GPT), a retail store on Union Street, and an AI radio station where Claude emerged as the best DJ. They developed a method to fork real environments into simulations mid-run, dramatically reducing simulation awareness. In a replay test of a Nazi song incident, Grok played it over 90% of the time, Gemini about half the time, while Opus and GPT refused every time.

Training Frontier Models to Out-Think Hackers — Uri Rolls, Arithmetic & Thom Wolf, Hugging Face

Jul 24, 2026 · 17:28

Uri Rolls of Arithmetic and Thom Wolf of Hugging Face argue that frontier models can be trained to out-think hackers, not just pattern-match vulnerabilities, by focusing on logic leaps in access control. Their benchmark, Mask Off, builds blackbox environments from real zero days found by human researchers, testing whether models can chain reconnaissance into exploitation. A live example: a Keycloak check validates admin by name while another checks by ID, so renaming oneself to the admin inherits privilege—GPT-5.5 and Opus probe everything but never make that logical leap. Results are brutal: only one solve at K1, with GPT-5.5 alone succeeding at K5. Rolls and Wolf argue this mirrors ARC-AGI’s challenge—models struggle to build dynamic world models—and that high-quality data and open source models can shift the economics of cyber defense, giving defenders a lasting speed advantage.

Notion's Token Town — Sarah Sachs, Notion

Jul 23, 2026 · 23:55

Sarah Sachs, Notion's AI engineering lead and contract negotiator, argues that AI companies must stop competing on token economics and instead build model-agnostic products that win on data flywheels, orchestration, and security. She advises treating every model supplier as a competitor, because frontier labs charge a markup on a markup for tokens they sell for first-party use. Notion's auto model routes 75% of traffic through a Switzerland-like system that swaps providers underneath, avoiding vendor lock-in. Sachs advocates routing by cost per capability per second, using open weight models for the moderate middle, and reaching for CPUs over GPUs (e.g., no LLM needed to turn a CSV into a PDF). She highlights the 'lethal trifecta' of private data, untrusted content, and external communication as the next security challenge, and demos Notion agents scoping a task, tagging teammates, and opening a PR. Her core message: optionality is leverage, and the product must transcend tokens.

Claude for Long-Horizon Tasks — Lance Martin, Anthropic

Jul 22, 2026 · 25:19

Lance Martin from Anthropic discusses how Claude's increasing task horizon enables asynchronous agents through decoupled architecture, verifier loops, and self-improving memory systems. He explains the shift from short task horizons (10–20 minutes) to 12+ hours, necessitating decoupling the brain (harness) from hands (sandboxes) for reliability and security. Verifier loops using separate contexts allow models to self-correct, demonstrated on the Parameter Golf benchmark with Opus 4.7. Memory systems inspired by human dreaming correct errors in-band, as shown in a Pokémon example where dreaming prevented repeated failures. Finally, org-level harnesses like Claude Tag provide shared identity and context for multiplayer proactive agents.

Don't Let the LLM Drive - Ornella Bahidika & Joel Allou, Microsoft

Jul 20, 2026 · 6:08

Ornella Bahidika and Joel Allou of Microsoft present their voice tutor Ace, arguing that the LLM should never control the flow in multi-step agents. They built Ace with a state machine that confines the model to narrow contracts per step, letting the harness validate outputs and decide next actions. This design allows them to use a cheaper, faster model (Haiku 4.5) instead of a heavy reasoning model (Opus 4.7 Cloud). They identify three decisions the LLM must never own: when the lesson ends, whether the student answered correctly, and what comes next. By engineering these checks outside the model, Ace achieves reliability in production, avoiding loops and early terminations that plague prompted-only approaches. The pattern applies to any flow agent, including coding agents, runbooks, and onboarding flows.

Your Voice Agent Doesn't Need a Frontier Model - Joel Allou & Ornella Bahidika, Microsoft

Jul 20, 2026 · 5:45

Joel Allou and Ornella Bahidika from Microsoft present ACE, an AI voice tutor that deliberately uses a small model instead of a frontier model, arguing that latency—not intelligence—is the critical constraint in voice applications. They demonstrate that a frontier model's reasoning time of multiple seconds breaks conversational flow, while their system keeps model latency under 950 milliseconds by extracting all logic and planning into a deterministic state machine. This scaffolding handles lesson progression, student mastery tracking, and response generation, leaving the small model (Haiku 4.5) with only the task of speaking. The result: a 900-millisecond response time that feels instantaneous. They acknowledge that scaffolding requires strict rules to prevent drift, but call it a one-time code investment that unlocks cost-effective, real-time performance. The episode’s core lesson: pick the fastest model your latency budget allows, then invest in external scaffolding to make it smart.

Stop Burning Tokens: Why self-improvement needs domain expertise first - Annabell Schäfer, Langfuse

Jul 18, 2026 · 17:39

Annabell Schäfer, Growth Engineer at Langfuse, argues that successful auto-improvement loops require domain expertise and high-signal target functions, not just generic evaluators. She details an experiment classifying arXiv papers with a minimal loop using GPT-5 for nano and Claude Opus 4.8 as an optimizer, achieving a 15% accuracy jump from 68% to 83% in four iterations. The first iteration alone gained 10% by adding classification rules and examples based on error analysis of a 200-item dataset. Schäfer advises replacing vague metrics like correctness with yes/no quality criteria (e.g., 'answer uses knowledge base'), and working with domain experts to identify failure modes and define what 'good' means. She emphasizes validation to prevent overfitting and treating the system as generalizing from representative examples, not just burning tokens on endless loops.

On AI and Knowledge — Pablo Castro, Distinguished Engineer & CVP for AI Knowledge, Microsoft

Jul 17, 2026 · 17:35

Pablo Castro, Distinguished Engineer and CVP for AI Knowledge at Microsoft, argues that AI agents require three knowledge types—intrinsic (model training), extrinsic (RAG and grounding), and learned (optimization loops)—and demonstrates how Microsoft's Foundry IQ, Azure AI Search, and Agent Optimizer deliver them. He shows that combining retrieval methods outperforms individual ones, citing Azure AI Search evaluations where combined methods improve evidence recall and answer completeness. He introduces Foundry IQ as a layered system with agentic retrieval that reflects on data to satisfy information needs before returning results. Castro demonstrates creating a knowledge base that connects unstructured data, structured Parquet tables, and the web, then linking it to an agent via MCP server. He also showcases the Agent Optimizer, which uses hill climbing on baseline evaluations to automatically refine agent instructions, tool definitions, and skills based on runtime traces, enabling continuous learning loops that capture organizational differentiation.

Imagination Engineering: "Live in the future and then build what's missing."

Jul 16, 2026 · 16:04

Eve Bouffard, Head of Design at Y Combinator, introduces 'Imagination Engineering' as the art of stretching the mind to invent what seems impossible, arguing that idea generation is the new bottleneck as AI models become incredibly capable. She shares her experiment in 'thinking in public' with a Slack channel called 'Eve Thoughts,' where she dumps her stream of consciousness, and then used Opus 4.8 to build a personalized website (EveBouffard.com) that aggregates and visualizes those thoughts. The site dynamically surfaces her projects, quotes, tools, and books, and even applies shaders and translations. She also built 'Shape of Minds'—a tool that analyzes commonalities across history's greatest minds, revealing patterns like taking naps and barely eating. Bouffard demonstrates how to spin up agents on demand for learning and productivity, emphasizing that anyone can now create software on the fly from their stream of consciousness.

Claude Fable, Claude Tag, and Anthropic's Culture — Cat Wu & Thariq Shihipar ft Simon Willison

Jul 15, 2026 · 51:30

This episode features Anthropic's Cat Wu and Thariq Shihipar discussing Claude Code, Claude Tag, and Claude Fable, arguing that these tools have shifted engineering from slow spec-driven processes to rapid, ambitious building. Thariq notes that with each model generation, delegation increased, and Claude Fable now enables one-shot features. Cat says engineers now need product taste over execution, as timelines shrink from six months to a week. Claude Tag, a proactive multiplayer agent, lands 65% of product engineering PRs internally by monitoring channels and remembering preferences. The team reduced Claude Code's system prompt by 80% for frontier models by removing examples and hard constraints, relying on model judgment. Auto mode, used internally since January, mitigates prompt injection through thousands of evals and Sonnet classifiers. Cultural hacks include default-public channels and a 'don't negotiate against ourselves' mindset, leading to ambitious builds like Thariq's Claude-powered video editing and a Street Fighter game.

The AI bugpocalypse is here. Now what? - Jack Cable, Corridor

Jul 12, 2026 · 19:44

Jack Cable, Co-Founder and CEO of Corridor, argues that frontier AI models can now find and exploit software vulnerabilities at scale, creating an 'AI bugpocalypse' that demands fundamental security changes. He cites 84% of developers using AI coding tools, with models like Mythos outperforming humans in finding vulnerabilities. He notes that basic vulnerability classes like buffer overflows remain prevalent despite decades of known fixes, and that memory-safe languages like Rust can eliminate 60-70% of such bugs, as shown by Android's drop from 75% memory safety vulnerabilities in 2019 to 30% in 2022. He warns that even advanced models introduce vulnerabilities 20-40% of the time (backsbench.com) and points to an Opus 4.6 incident that lost $2M. His recommendations to Congress include preventing vulnerabilities in new code, hardening open-source software, and fostering American open-weight models, arguing that defenders need access to these models to stay ahead.

Design Patterns for AI Trust: Juries, Libraries, and Agent Tiers — Alex Bauer, Upside.tech

Jul 11, 2026 · 17:09

Alex Bauer, co-founder of Upside, argues that go-to-market teams can solve AI's trust problem by managing agents like humans, using patterns such as a librarian for just-in-time knowledge and a jury-and-judge model for subjective decisions. He demonstrates three solutions: first, scaffolding an AI website rebuild with anchor assets like a product capabilities reference that ladders directly into homepage copy; second, a librarian that consults documentation and prior failed queries before answering questions like "How much pipeline in Q1?" to catch fiscal-year or stage definitions; third, a jury-and-judge workflow for multi-touch attribution where independent analysts research deals and a judge weighs reasoning quality before producing a consensus verdict. Bauer warns against low-intelligence models for important work, citing Slackbot's MCP integration as "horrifically stupid," and advocates using tier-two platforms with sub-agents, plan mode, and full MCP support.

Everything we knew about software has changed — Theo Browne, @t3dotgg ​

Jul 8, 2026 · 16:02

Theo Browne argues that AI model evolution—from Sonnet 3.5’s tool-calling to Opus 4.5's long-running tasks and Mythos's orchestration—requires engineers to think bigger and wider. He compares current developer habits to skeuomorphism in iOS 7, urging rejection of legacy constraints like Git's inability to commit environment files and terminal-centric workflows. Browne introduces a shifted tier system: what was a startup is now a side project; a Markdown file running on a cron job can replace a company's product. His own PR triage service became a Markdown file updated daily via cron. He advocates building breadth over depth—architecting products so users can extend features, enabling small teams to compete with AWS or Salesforce. 'If your idea doesn't feel stupid, it's because your idea is not big enough,' he concludes.

Building an ACP-Compatible Agent Live — Bennet Fenner, Zed

Jul 8, 2026 · 18:19

Bennet Fenner from Zed demonstrates building an ACP-compatible coding agent live, showing how the Agent-Client Protocol lets agents and editors communicate over JSON-RPC. He starts with a minimal agent that reads and edits files, then adds ACP support using the TypeScript SDK—implementing initialize, session creation, and prompt handling. Next, he streams model tokens via session updates, emits tool call progress with file paths and diffs, and proxies file system operations through the client. The demo culminates in the agent bootstrapping itself to add a terminal tool, which then runs shell commands inside Zed. ACP currently works over stdio; a remote transport is in development.

What if the harness mattered more than the model? - Aditya Bhargava, Etsy

Jul 7, 2026 · 32:04

Aditya Bhargava, a Staff Engineer at Etsy, argues that the harness—the tooling around a model—matters more than the model itself, especially for weaker models, and introduces Agency, a language he built for constructing agents. He cites a benchmark where only the harness changed, producing a 52.4% to 76.2% score range (over 20 points), showing that a better harness can make local open-source models perform near proprietary ones. Through seven evolving examples of a coding agent, he demonstrates key harness improvements: adding tools, ensuring safety via handlers and partial function application (PFA) to lock directory access, implementing a reason-act-feedback loop, using subagents to add capabilities without context bloat, and self-optimization via a built-in optimizer to systematically measure and improve prompts. He advocates for building expertise in harnesses to reduce reliance on proprietary models, enabling use of locally runnable ones. Agency provides simple tool syntax, safety features like interrupts and true pause/resume, and built-in optimizers for measured improvement.

Field Guide to Fable — Thariq Shihipar, Anthropic

Jul 6, 2026 · 19:28

Thariq Shihipar of Anthropic discusses Fable and argues models improve in "spiky" ways: a chat model fails to list Pokémon ending in "aw" (Croconaw and Dreadnaw), but Claude Code fetches and filters the list in seconds – a gap he calls "capability overhang." He explains that to unlock Fable, Claude Code cut 80% of its system prompt because heavy instructions now constrain a more imaginative model, and the "ask user question" tool evolved from barely working under Opus 4 to generating embedded HTML questionnaires. He shares techniques like blind-spot passes and interviews to surface unknown unknowns, and reflects on the grief of moving from hand-coded programming to agentic workflows. Shihipar urges engineers to reject trade-offs – "good, fast, cheap: pick three" – and instead demand all three, citing a four-hour keynote deck built with Fable as proof that agents can deliver ambitious work faster.

We Cut 94% of AI Coding Tokens With a Local Code Index - Rajkumar Sakthivel, Tesco

Jun 28, 2026 · 10:43

Rajkumar Sakthivel and his friend Faz built Code Context Engine (CCE) to cut AI coding costs after their bill jumped from £15 to £200 in one month. They found that 45,000 tokens were sent per query but only 5,000 were needed, with 90% of cost being input context. Their solution is a local retrieval layer that uses AST-aware chunks, combined vector and keyword search, and a weighted scoring heuristic (50% meaning, 30% keyword, 20% recency) that runs in 0.4 milliseconds. On a FastAPI test, tokens per question dropped from 83k to 4.9k—a 94% reduction—with 90% accuracy. They emphasize that fixing the input, not the model, yields the biggest savings, and their open-source tool shares a single index across Claude Code, Cursor, Copilot, and Codex.

Recursive Coding Agents - Raymond Weitekamp, OpenProse

Jun 25, 2026 · 23:48

Recursive Language Models (RLMs) are a new test-time compute paradigm that unify reasoning and tool calling, and Raymond Weitekamp of OpenProse argues they can turn coding agents into reliable, recursive problem solvers. He demonstrates that RLMs can process millions of tokens beyond context windows, and a small Qwen 3.5 9B model using RLMs beats GPT-5.4 and Opus on the LongCOT benchmark. Weitekamp shows how Claude Code's Dynamic Workflows now make it an RLM, and how OpenProse allows any coding agent to be recursively orchestrated with declared sub-agent work and dependencies. Use cases include repo-scale migrations, deep research over file systems, and adversarial red-teaming audits. He concludes that RLMs represent the next paradigm of inference-time compute, turning mismanaged genius agents into trustworthy outcomes by capturing golden sessions as reusable workflows.

Why Can't Anyone Answer Questions About the Business? — Garrett Galow, WorkOS

Jun 11, 2026 · 19:06

Garrett Galow from WorkOS built Studio, an internal workspace where anyone can ask natural language questions against Snowflake, Linear, and Notion, and get reusable widgets instead of filing a request. The LLM generates declarative JavaScript widgets that call data sources directly, making subsequent runs deterministic and cheap. Three techniques made it reliable: preflight sequencing injects schema context only when a tool is invoked, a layering rule tells the model to distrust its own knowledge about WorkOS and use primary sources, and query validation catches valid SQL that returns zero rows before hardcoding it into a widget.

Why More Context Makes Your Agent Dumber and What to Do About It — Nupur Sharma, Qodo

Jun 8, 2026 · 26:27

Nupur Sharma from Qodo explains why giving an AI agent more context often makes it dumber, describing the 'U curve' where models attend only to the start and end of inputs while dropping the middle. She covers practical fixes: iterative retrieval, hierarchical summarization, and self-correction with honest cost tradeoffs. The talk introduces the 'orchestration paradox'—smart models waste most tokens figuring out how to solve a problem rather than solving it—and Qodo's 80/20 hybrid: high-reasoning models for open-ended discovery, lighter deterministic models for validation. Sharma walks through Qodo's code review architecture: a context collector feeds specialized agents (security, code quality, etc.), a judge node recombines results against PR history, and every accepted or rejected suggestion shifts weighting for future reviews.

Beyond Components: Designing Generative UI for MCP Apps — Ruben Casas, Postman

Jun 3, 2026 · 16:58

Ruben Casas from Postman argues that AI models can now write better frontend code than he can—his prompt to rewrite his blog produced a search box with blur animation and accessibility out of the box—yet most agent UIs still invoke static prebuilt components. He presents three levels of UI generation: static components (AG UI, Goose) passing props to predefined React elements; declarative UI where the model generates JSON or YAML for a rendering engine (e.g., Vercel's JSON Render), which he deems the current ideal balance; and fully generative UI where the model writes HTML, CSS, and JavaScript on demand, as in his weather agent that does so in one tool call. The key obstacle is trust, necessitating sandboxing, and MCP apps with their double iframe default are the best delivery mechanism. He likens today to early TV—radio shows with cameras—and predicts the future lies beyond components in collaborative human-agent interfaces on shared canvases, as seen with the Skeletro MCP app.

What if the network was the sandbox? — Remy Guercio, Tailscale

Jun 1, 2026 · 24:29

Remy Guercio from Tailscale argues that standard sandboxing conflates execution isolation with access control, proposing Aperture—an LLM gateway built on Tailscale's WireGuard identity network—which gives every connection verified identity (user, tag, or group) so agents get placeholders instead of real API keys, making exfiltration impossible. Aperture provides visibility into every tool call, bash command, and MCP request without instrumentation inside the container; internally at Tailscale, bash dominates over structured tool calls. Access permissions are configured via Tailscale's grants and ACLs, supporting quotas, cost controls across providers, and webhooks for tool calls. The gateway works at the LLM layer, capturing even non-tool-call agent behaviors like direct code execution, and is available on Tailscale's free plan.

The AI Skill I Rely On Daily — Priscila Andre de Oliveira, Sentry

May 27, 2026 · 17:05

Priscila Andre de Oliveira, a senior software engineer at Sentry, reveals that 67% of her AI usage is for comprehension and only 2% for code generation, based on analyzing 116 of her own Claude sessions. Working in Sentry's 15-year-old codebase with 100 PRs merged daily and 100,000 organizations depending on it, she built a personal skill called 'Catch Me Up' with six exploration modes (architecture, conventions, feature traces, syntax, testing, history). She argues that understanding what the agent found before letting it plan and implement prevents misaligned mental models that produce slop code. The episode emphasizes that in large, complex codebases, AI's biggest unlock is comprehension, not generation, and advises developers to align their mental models before prompting.

Run Frontier AI at Home — Alex Cheema, EXO Labs

May 26, 2026 · 1:45:02

Alex Cheema of EXO Labs argues that running frontier AI locally has 100x improvement potential in cost and performance, demonstrated with GLM 5.1—a trillion-parameter model—running across four Mac Studios at roughly 20 tokens per second for $40,000. He details kernel fusion that recovered 30% performance on Qwen 3.5 by eliminating unnecessary kernel launches, and RDMA integration that cut node-to-node latency from 300 microseconds to single digits, enabling tensor parallelism to actually scale. Cheema advocates splitting inference: prefill on compute-dense hardware (e.g., an RTX Spark) and decode on high-bandwidth hardware (e.g., Mac), cutting large-prompt inference roughly in half. He warns against misleading benchmarks like one-bit quantized models, and outlines how multi-agent setups, test-time scaling, and continual learning could further improve local inference efficiency. The talk includes a live demo of GLM 5.1 across four Mac Studios connected via Thunderbolt 5, and a preview of EXO's upcoming benchmarking site to track intelligence per Joule.

Fast Models Need Slow Developers — Sarah Chieng, Cerebras

May 22, 2026 · 18:02

Sarah Chieng from Cerebras argues that the 20x speed increase of models like Codex Spark (1,200 tokens/sec vs. 40-60 for Sonnet/Opus) forces developers to rethink workflows, or risk generating technical debt at unprecedented scale. She presents a practical playbook: validation and linting become free at every step, so must run continuously; developers can generate 75 component variations across five sub-agents and cherry-pick the best; and with context filling in 30 seconds instead of ten minutes, a four-file external memory system (agents, plan, progress, verify) maintains continuity between sessions. Chieng emphasizes real-time collaboration with the model rather than spawning agents and walking away, and advocates using a slower planner model with a fast executor to orchestrate agents effectively.

Don't Build Slop (4 Levels of AI Agent Maturity) - Ara Khan, Cline

May 19, 2026 · 18:52

Ara Khan of Cline presents four levels of AI agent maturity, arguing that most builders suffer from mass psychosis and should focus on thoughtful architecture. Level one is using existing frameworks for rapid prototyping. Level two introduces five rules: treat every agent as a state machine, keep system prompts minimal (e.g., GPT-5.3's prompt is one-third the size of GPT-5's to avoid sensory overload), integrate with a CLI for pseudo-RL pipelines, avoid building slop by designing architecture manually, and master frontier model APIs to leverage reasoning traces correctly. Level three advocates Kanban boards as the ideal UX form factor for managing parallel inference-bound agents, a prediction he made on March 26 that Claude Code shipped ten hours before this talk. Level four involves shipping agents to the cloud for scalability, enabling long-running tasks and parallel execution without local dependencies.

Give Your Agent a Computer — Nico Albanese, Vercel

May 12, 2026 · 1:08:53

Nico Albanese demonstrates building an agent with Vercel's AI SDK v6, centered on the insight that giving an agent a file system transforms its behavior—it follows through on long tasks, stays on track, and builds on prior work. He walks through creating a tool loop agent from scratch, adding provider-executed web search with typed UI components, and integrating Vercel's persistent named sandboxes that snapshot state after inactivity. The agent gains a bash tool for file system access, a memories.md file for persistent memory injected into instructions each turn, and instructions to generate Python scripts for repeatable tasks so it accumulates tools across sessions. Albanese explains how these patterns scale: an internal agent called D0 reduced customer support tickets by 90% with a 95% satisfaction rate, and his personal coding agent ran for 104 minutes in one turn, using 316 tool calls and only 32% of GPT-5.4's context window with zero compaction. The session concludes with a preview of a larger sub-agent system that uses durable workflow steps and background sub-agents to keep the main thread under 7,000 tokens.

Viktor: AI Coworker That Lives in Slack — Fryderyk Wiatrowski

May 11, 2026 · 19:30

Fryderyk Wiatrowski, co-founder of Viktor, explains how his AI employee lives in Slack—no web UI—participating in channels and threads like a teammate, inheriting integrations from whoever connected them first, and handling tasks that take ten minutes. He details the challenges of scaling a personal agent to a company agent: memory management across hundreds of users, managing Slack's complex input surface (threads, DMs, edits, emoji reactions), and preventing context leakage between channels. He shares that swapping the underlying model from Opus to GPT-5.4 caused user backlash due to personality differences, and describes the need to earn proactivity trust to avoid security alarms. Viktor's advantage is shared context—only one person needs to connect an integration for the whole team—but warns against giving it personal email access, as illustrated by a customer story. The episode argues that a great AI coworker requires helping get work done, knowing the company context, and being friendly.

Vibe Engineering Effect Apps — Michael Arnaldi, Effectful

May 7, 2026 · 1:43:04

Michael Arnaldi demonstrates that cloning the Effect library's repository into a project, rather than relying on prompts, is the most effective way to give coding agents the context they need to build reliably with Effect. Starting from an empty repo, he sets up a Bun, Vitest, and TypeScript project, adds the Effect repo as a git subtree, and creates an agents.md to guide the agent. He then uses the agent to research patterns, resulting in a fully functional Todo HTTP API with OpenAPI docs, SQLite persistence, and tests — all from scratch in under two hours. The workshop stresses that treating the library code as part of the project, combined with strict diagnostics and pattern files, makes agents effective even in unfamiliar codebases. Arnaldi also discusses the importance of workflow systems like Effect Cluster for long-running AI processes, where server failures become likely.

Replacing 12K LoC with a 200 LoC Skill — David Gomes, Cursor

Apr 30, 2026 · 19:22

David Gomes shows how Cursor replaced 12,000 lines of code for Git WorkTrees and best-of-en features with roughly 200 lines of Markdown using agent skills and subagents. He explains the original implementation's complexity (15,000 lines deleted) and the new slash commands: /worktree, /best-of-en, /apply, /delete. Pros include less maintenance, ability to switch mid-chat, multi-repo support, and better judging with the parent agent stitching results. Cons: models sometimes forget to stay in the WorkTree over long sessions, perceived slowness, and reduced discoverability. He details future improvements through evals and RL training, plus a native WorkTrees implementation in Cursor 3.0 and exploration of non-Git parallelization primitives.

Collaborative AI Engineering: One Dev, Two Dozen Agents, Zero Alignment — Maggie Appleton, GitHub

Apr 26, 2026 · 17:43

Maggie Appleton, a staff research engineer at GitHub Next, argues that the current paradigm of one developer with a fleet of solo agents leads to zero team alignment, making software a team sport that requires shared context. She presents ACE (Agent Collaboration Environment), a multiplayer research prototype that provides shared chat sessions backed by micro VMs and Git branches, enabling real-time collaboration where designers, PMs, and developers can all prompt agents together. The talk emphasizes that implementation is now cheap and fast, so the hard question is "should we build it?" — requiring early, constant alignment to avoid wasted work, coordination debt, and unrequested features. ACE aims to reclaim time for rigorous critical thinking by letting teams plan, build, and review in a single shared space, with proactive agents summarizing team activity and unfinished work. Appleton contends that quality becomes the differentiator in a world of cheap code, and tools like ACE can help teams build a few exceptional things rather than a thousand crappy ones.

Full Walkthrough: Workflow for AI Coding — Matt Pocock

Apr 24, 2026 · 1:36:30

Matt Pocock presents a hands-on workshop on building a full AI-assisted coding workflow, arguing that software engineering fundamentals—not hype—make agents effective. He introduces the 'smart zone' and 'dumb zone' of LLMs (performance drops after ~100k tokens) and the 'Memento problem' (agents forget between sessions). His process starts with a 'Grill Me' skill that relentlessly questions the user until shared understanding is reached, then produces a PRD without reading it, slices work into vertical 'tracer bullet' issues, and runs agents AFK using TDD. He advocates designing codebases with deep, testable modules and shows Sandcastle, a TypeScript library for parallel agent execution with separate implementer (Sonnet) and reviewer (Opus). The workshop transforms ambiguous briefs into shippable features while keeping humans in the loop for QA and taste.

How AI is changing Software Engineering: A Conversation with Gergely Orosz, @pragmaticengineer

Apr 21, 2026 · 26:42

Gergely Orosz explains how AI is reshaping software engineering, from token maxing at Meta, Microsoft, and Salesforce—where engineers use AI tools to inflate token counts out of fear of layoffs and performance reviews—to the broader shift in the engineer's role toward orchestrating agents rather than managing people. He notes that big tech companies like Uber, Airbnb, and Shopify are building custom internal AI infra (MCP gateways, coding agents) to stay ahead, even tolerating high churn for a six-month competitive edge. Shopify secured GitHub Copilot a year early by offering feedback from 3,000 engineers. Gergely also shares how The Pragmatic Engineer hit product-market fit: 100 paid subscribers before publishing, reaching 1,000 in six weeks, and becoming the #1 paid tech newsletter by focusing on two deep-dive articles per week for two years.

Full Workshop: Build Your Own Deep Research Agents - Louis-François Bouchard, Paul Iusztin, Samridhi

Apr 20, 2026 · 1:57:03

Louis-François Bouchard, Samridhi Vaid, and Paul Iusztin built a combined system of a deep research agent and a deterministic writing workflow to automate high-quality technical content creation, trading flexibility for control. The research agent uses an MCP server (FastMCP) with tools for Google search, YouTube analysis, and report compilation, powered by Gemini and orchestrated by Claude Code. The writing workflow is a static pipeline that loads a user guideline, writing profiles (structure, terminology, character), and few-shot examples into a system prompt, then applies an evaluator-optimizer loop: a reviewer outputs structured pydantic objects with profile/location/comment to flag violations, and the editor applies them in priority order (guideline > research > profile). They stress that research demands autonomy while writing needs constraint, and show how to avoid AI slop by banning specific words and using profiles. For observability, they use Opik to capture all traces, and for evals they built a 20-sample dataset from real LinkedIn posts, split into train/dev/test, to calibrate an LLM judge (binary pass/fail) that they run on dev and test splits to measure F1 and detect…

Code Mode: Let the Code do the Talking - Sunil Pai, Cloudflare

Apr 19, 2026 · 19:40

Sunil Pai of Cloudflare presents 'Code Mode,' a paradigm where AI agents generate executable JavaScript instead of using JSON-based tool calls, drastically reducing token usage from 1.2 million to 1,000 for Cloudflare's 2,600 API endpoints. He demonstrates a live (if glitchy) demo of the Mythical server handling a DDoS response in one shot, and shares the 'harness' architecture: a sandboxed, capability-secure environment that grants only explicit APIs. The talk explores emergent behaviors like Kenton Varda's canvas where the model played tic-tac-toe by inspecting raw stroke state, and envisions long-running workflows, generative UI, and a future where code—not buttons—defines user interaction. Pai argues that developers must optimize for agent-facing DX with searchable docs, clear errors, and capability-based security, as the next billion users will be code-generating bots.

System Design for Next-Gen Frontier Models — Dylan Patel, SemiAnalysis

Feb 11, 2025 · 18:29

Dylan Patel of SemiAnalysis breaks down the inference challenges for next-generation frontier models like GPT-4 (1.8 trillion parameters) and upcoming models trained on 100,000+ GPU clusters. He emphasizes that prefill (prompt processing) is compute-intensive while decode (token generation) is memory bandwidth-intensive, creating a systems problem where serving 64 users at 30 tokens/second requires 60 terabytes/second of memory bandwidth. Patel details engineering strategies such as continuous batching to improve batch utilization by 10-100x, disaggregated prefill to isolate noisy neighbors and maintain time-to-first-token SLAs, and context caching (like Google's) to cache KV cache on CPU/storage instead of GPU memory, dramatically reducing prefill costs. He warns that open-source tools like LLaMA.cpp lack these optimizations, making high-performance serving of models like LLaMA 405b infeasible without libraries like vLLM or TensorRT-LLM. On scaling, Patel notes that 100,000 GPU clusters (e.g., Microsoft's Arizona data center consuming 150 MW) face reliability issues — optical transceivers fail every five minutes — and straggler chips (silicon lottery) can degrade training…

Building an AI assistant that makes phone calls [Convex Workshop]

Feb 9, 2025 · 50:53

Tom Redmond, head of DX at Convex, builds Floyd, an AI phone-calling assistant using Convex, GPT-4, Google Cloud speech-to-text, and Twilio. Floyd lets users make voice requests (e.g., calling a school to report a sick child) via a web app; transcription, context retrieval, and real-time conversation with a human on the other end happen through a reactive database architecture. Redmond explains how Convex's WebSocket-based infrastructure solves serverless latency issues, enabling live transcript streaming and status updates without polling. He demonstrates two calls (school absence and flower order), noting that latency (2–5 seconds) comes from OpenAI text-to-speech and growing prompts, with plans to use OpenAI Threads and faster TTS services. The prototype explicitly discloses it's an AI to avoid misrepresentation, and Redmond envisions AI agents efficiently handling service bookings.

Lessons from the Trenches: Building LLM Evals That Work IRL: Aparna Dhinkaran

Feb 6, 2025 · 18:49

Aparna Dhinakaran, co-founder of Arize AI, distinguishes between model evals (e.g., Hugging Face leaderboard) and task evals for real-world LLM systems, arguing that production applications need component-level evaluations like router and parameter evals. She demonstrates a chat-to-purchase app where a router function call misidentifies user intent, showing how Phoenix open source tool traces errors and provides explanations to iterate. Dhinakaran advises using categorical over numeric LLM-as-judge scores because numeric outputs tend to be binary (0 or 10) and lack granularity. Presenting needle-in-haystack research, she notes GPT-4 struggles retrieving facts placed early in large context windows, and in retrieval-with-generation tasks, Anthropic’s Claude 2.1 outperforms GPT-4 due to verbose reasoning, a gap closed by prompting GPT-4 to explain itself first.