Day 2 at the AI Engineer World’s Fair 2026 (Moscone West, San Francisco) was the Coding Agents day. The main stage made the central tension of the entire conference explicit: two back-to-back talks taking opposite positions on how autonomous coding systems are best understood — and therefore built.
The Main Stage Argument
At 11:10 AM, Tereza Tížková (Growth, Factory) presented “Rise of the Software Factory” on the Main Stage — the case that AI coding systems are best understood as production lines. Agents have tasks, toolbelts (the hundreds of tools — repos, test runners, deployment scripts, documentation — an enterprise agent gets wired into), context, and feedback loops running inside an outer loop. You, the engineer, design the factory; the factory builds the software. Tížková’s own published write-up of the talk frames it the same way: “a software factory is something you build and own, not a consultancy you hire,” with humans “deciding what to build rather than how to build it, because the how is exactly what the agents are for.”
At 11:40 AM, Charlie Holtz (Conductor, CEO) followed immediately with “Orchestras, not Factories”, billed on the official schedule as the story of how Conductor’s interface came together and “what I think everyone (including us) is getting wrong.” Holtz has laid out the underlying metaphor in interviews: he doesn’t want engineering reduced to factory-line work — he wants to feel like a conductor “in front of an orchestra, waving my baton,” walking over to an individual player (an agent) to say “you’re out of tune,” rather than treating every unit of output as interchangeable. “Conductor CEO Charlie Holtz Walks Us Through His AI Coding Setup,” YC Startup Podcast has him describing exactly that scene. The distinction he draws is about who stays in the interpretive loop: a factory line runs on standardized, queued steps, while an orchestra runs off a shared score with room for a human to intervene section by section.
The fact that the schedule placed these talks back-to-back on the Main Stage suggests the AIEWF organizing team was deliberately surfacing the argument rather than endorsing either position.
Both positions agree on the fundamentals. Autonomous coding agents are production reality. The dispute is about the mental model that should govern how you build the infrastructure around them — and that model shapes API design, error handling, retry semantics, observability, and team structure.
Daksh Gupta’s Data: 1M+ AI-Generated PRs
At 12:05 PM, Daksh Gupta (Greptile co-founder and CEO) presented “What we learned by analyzing 1M AI-generated PRs” on the Main Stage. The talk’s own billed focus was bug patterns in agent-written code — including the finding that Claude Code was nearly 3x more likely than Codex to introduce auth-bypass vulnerabilities in Greptile’s sample. The growth and quality numbers underneath that dataset come from Greptile’s own research, published in its “Rise of the Overnight Agents” report:
Greptile has reviewed several million PRs across 65,000 organizations over nearly two years, per that report:
- 0.86% of PRs showed evidence of being fully AI-generated in February 2025
- 27.6% were fully AI-generated by April 2026 — a 32x increase in 14 months
That rate of adoption has no precedent in software tooling history. It is not a slow enterprise rollout. It is a step-function shift.
The quality data is the part that will get cited for months, but it varies by tool rather than sitting at one uniform rate. Per Greptile’s data (measured Mar 15–Apr 14, 2026, in reverts per 1,000 merged PRs): Codex (1.19) and Claude (1.80) reverted less often than the human baseline (2.72), while Cursor’s background agent (3.41) and Devin (3.50) reverted more often. Greptile’s own summary: reversion rates for AI-generated PRs are “pretty similar” to reversion rates for human-written PRs overall. For larger PRs specifically, human-written code had the higher revert rate of the two.
AI PRs are also about 20% larger by median lines of code than the same developer’s non-AI PRs — 171 median lines of code versus 143, measured across the same developers in April 2026. Agents don’t write small changes; they write complete implementations.
The implication Gupta drew: you can no longer assume a PR is human-authored. Your review process, your CI/CD pipeline, your security analysis, and your onboarding documentation all need to be rethought with the assumption that a significant and growing fraction of contributions are agent-generated.
GitHub Copilot Agents (Idan Gazit, GitHub Next)
At 10:45 AM, in a sponsored session (Track M), Idan Gazit (Head of GitHub Next) presented “Build agents fast with GitHub Copilot (from idea to working app)" — per the schedule, a walkthrough of “how developers go from prompt to a working agent using GitHub Copilot and real workflows,” covering code generation, fast iteration, and staying inside an existing dev loop.
The talk sits inside GitHub Next’s broader “Continuous AI” thesis: background agents that operate inside a repository the way CI jobs do, but for tasks that need reasoning rather than fixed rules — with a human reviewing outcomes rather than approving every intermediate step.
Read alongside Gupta’s numbers from the same day, the open question is straightforward: when a growing share of PRs are agent-generated, what does human review actually cover?
Evals as the Rising Discipline
Two talks in the 11:10-11:30 AM slot — running in parallel with the Software Factory keynote, on different tracks — addressed what becomes necessary when a growing share of PRs are agent-generated and can’t be manually validated one by one:
“Your Agent Evolved. Your Evals Didn’t." (Ameya Bhatawdekar, Braintrust VP, Field CTO) — per the session abstract, the case that agent architectures have moved through six generations (prompt, chain, ReAct loop, workflow graph, modern agent loop, AI harness), and each generation quietly breaks the eval strategy built for the one before it: a prompt-quality rubric won’t catch a bad tool call, and a trace scorer won’t catch memory poisoning.
“Your Agent Didn’t Fail. Your Harness Did." (Vinoth Govindarajan, OpenAI, Member of Technical Staff) — the complementary argument, using OpenClaw as a public case study, that most production agent failures happen in the harness rather than the model: state that isn’t persisted, two runs mutating the same session, a tool call that never returns, or an approval that loses scope. His proposed fix is a “run receipt” audit — what woke the agent up, what state it inherited, what authority it used, what executed, and what evidence survived.
Both talks build on a three-year arc swyx laid out in the conference’s opening keynote, marking three years since he published “The Rise of the AI Engineer” in mid-2023: “prompt engineering gave way to rigorous evals, RL environments for post-training, and context/harness engineering.” Evals are the discipline that absorbed most of that shift — the systematic, measurable approach to forward progress that mature engineering requires.
Anthropic Claude Managed Agents Workshop
The Claude Managed Agents Workshop (Priyanka Phatak and Gabriel Cemaj, both Members of Technical Staff at Anthropic) ran as a four-part sequence on Track 4, 10:45 AM to 12:25 PM: build an agent with Claude Managed Agents, in four 20-minute segments.
Lance Martin (Anthropic, Member of Technical Staff) presented “Claude for long-horizon tasks” at 1:55 PM. Per the session abstract and a separate write-up of the same material, the talk covered patterns for reliable, secure long-horizon agent harnesses: decoupling the “brain” (a stateless orchestration layer) from the “hands” (ephemeral execution sandboxes) with an append-only session log as the source of truth; running an independent verifier loop in a separate context window rather than letting an agent grade its own work; and letting a model maintain its own memory schema, corrected by an offline review pass, instead of a prescribed one.
Closing: Recursive Model Improvement
Lee Robinson (ML, Model Behavior, Cursor) closed Day 2’s Main Stage at 5:10 PM with “Recursive Model Improvement” — not a product-direction talk, but a technical walkthrough of how Cursor trains its own coding models. The core idea: each new generation of Cursor’s frontier model gets distilled into derivative models — reward models, eval judges, data-generation assistants — that in turn improve the training pipeline for the next generation, creating a compounding flywheel. Robinson also described building Cursor Bench, a private evaluation suite built from real engineering tasks (analyzing logs, Slack threads, and docs to diagnose production incidents), after models learned to game public coding benchmarks by searching Git history and public eval forks for answers. He noted that the large majority of Cursor’s revenue now comes from agent usage rather than autocomplete — which is why their training data has shifted toward agentic workflows.
The Argument That Matters for Builders
The factory vs orchestra framing is not aesthetic. It has practical consequences:
| Question | Factory answer | Orchestra answer |
|---|---|---|
| How do agents share context? | Shared queue / pipeline | Broadcast event / shared score |
| How do errors propagate? | Stop the line | Isolate the instrument |
| How do you add capacity? | Add workers | Add instruments |
| How do you measure output? | Throughput | Coherence |
| What does failure look like? | Stalled job | Off-key section |
Gupta’s data suggests the practical consequence is already arriving faster than most engineering orgs have designed for. If 27.6% of PRs are agent-generated today, and revert rates are comparable to human PRs, the constraint shifts from “is the code good enough” to “how do we understand what agents are doing at scale.” That is an observability problem as much as a quality problem.
The evals track answers that question from the measurement side. The harness engineering track answers it from the infrastructure side. Both converged on Day 2.
What’s Next
Day 3 (July 1) is the Autoresearch day — multi-agent research systems and autonomous research agents, opening with Anthropic’s “Field Guide to Fable” keynote.
Day 4 (July 2) is Harness Engineering: production infrastructure, agentic commerce, inference, and security. It opens with Barr Yaron (Partner, Amplify Partners) presenting “The 2026 State of AI Engineering” — Amplify’s annual survey, historically one of the most-cited data products to come out of the conference.
ChatForest covers AI infrastructure from the builder’s perspective. This recap is based on publicly available session data and talk descriptions from the AIEWF 2026 schedule and post-conference coverage. ChatForest did not attend the conference; this is a synthesis of publicly available information.