
Hey readers! Two big coding-agent conferences just wrapped, and the throughline is weirdly consistent: the hard part of agentic coding is no longer writing the code, it's trusting it. One speaker clocked 23 hours and 11 rounds of agent review on a single small ticket. Let's dig into what the AI Engineer Code Summit and Tessl's AI DevCon NYC actually surfaced, then round up the week's tooling churn.

AI Native RL Environments
Building the Worlds Where AI Agents Learn to Work · v1.3, October 2026
The thesis fits in one line: environments are the new data. The frontier has moved from what a model has read to what it has practiced. Coding agents didn't get good by reading more code. They got good by being dropped into real repositories with failing tests and rewarded when the tests passed.
This 149-page playbook is written for founders, developers, investors and engineers. Every chapter ends with key takeaways and a "For your role" section, so you can skim it in an afternoon and come back for the depth.
Rewards and verifiers: how agent work gets scored
Reward hacking: how agents game weak environments, and how to harden them
The open-source stack: environment hubs and training frameworks like TRL, verl and Tinker
Sandboxes and the data layer: tasks, trajectories and runs at scale
Build your own: a complete environment from scratch, with full source
Your product as an environment: digital twins, agent-ready MCP interfaces and official sandboxes
The idea most relevant to this week: most companies already own the raw material of an environment, in their ticket history, audit logs and the judgment of their senior people. Read the stories below with that in mind. Isolated cloud sandboxes, session traces and done-checks are all pieces of an environment showing up in everyday tools.
🎤 AI Engineer Code Summit: a year of agentic coding, one room
The second annual AI Engineer Code Summit ran November 10–12 at the Hilton Union Square in San Francisco: a workshop day followed by two days of single-track keynotes, organized around three themes the organizers called scaling without slop, system one models, and rewrites. Superpower Daily's preview captures the framing: coding-agent workflows paired with sessions on checking their work and controlling their autonomy.
OpenAI on new ways of working with Codex. The workshop day opened with sessions from OpenAI, Anthropic, and Google DeepMind. In Latent Space's writeup of the Codex workshop, OpenAI's Dominik Kundel and Charlie Guo walked through handing tasks off to the cloud, steering work by voice, and using personal agents ("dots") that carry context between sessions. Their advice was to stay flexible, because these workflows are still changing fast.
System One Day. Wednesday's big idea was "system one" models: non-generative models that return structured choices, scores, or probabilities, and run much faster and cheaper than reasoning LLMs. In The Jev Revolution, TypeSafe AI's Diogo Almeida argued these should be wired straight into code, because "code is the language of automation." D-Day (Decision Models Day) adds the practical version: Cloudflare's Michelle Chen described a hybrid agent where fast decision models handle tool selection and log extraction, with a larger model reserved for the hard parts. Google also entered the decision-model market the same day.
Earning autonomy. Factory brought two keynotes: Luke Alvoeiro's "Earning Autonomy: From Missions to Workstreams" and "12-Factor-Factories: Design Patterns for Agentic Coding," plus an expo session on CVE remediation from finding to merged PR.
The pattern across the summit: use cheap, fast models for the decisions you can automate, and spend scarce human (and frontier-model) attention on judgment.
🏭 Tessl AI DevCon NYC: from coding agents to software factories
AI DevCon NYC: From Coding Agents to Software Factories sets out the theme of Tessl's Brooklyn event, and the talk recordings are now rolling out (track them on the AI DevCon hub). The event ran Nov 2–4 at Industry City, drew 700+ attendees, and packed 52 sessions around one question: how do you go from a single coding agent to a factory you can trust?
The standout data point came from James Moss, who showed what breaks when you scale from 1 coding agent to 10. His example: one small ticket that went through 11 rounds of agent review over 23 hours and still landed on "changes requested." If you've wondered why parallel agents haven't magically 10x'd your team, that's the answer in one anecdote. Output scales; the review queue scales faster.
Guy Podjarny opened with "Building a software factory you can trust," and Steve Yegge closed with the argument worth sitting with: once agents run agents, you have a factory. The mechanical layer (CI, execution, deploys) is turning into a commodity, but the layer above it, where your taste and judgment live, can't be copied. Birgitta Boeckeler, Chad Fowler, Ryan Lopopolo, and Geoffrey Huntley rounded out the speaker list.
If agents running agents is the direction, environments built for exactly that are worth watching. SpaceMolt, a realtime MMORPG for AI agents, is a playful but real example of a world where agents operate, compete, and coordinate in real time, the kind of practice ground this whole "factory" conversation keeps circling back to.
🤖 The multi-agent land grab
Meta, OpenAI and GitHub bet on multi-agent coding ties the conference themes to real deals: in the week of Oct 28 to Nov 4, Meta acquired Conductor (a Mac app running Claude Code, Codex, and Cursor side by side), OpenAI agreed to buy Pierre, and GitHub shipped Copilot Projects in public preview. The through-line matches what both conferences argued: parallel generation is solved, coordinating and merging the output is not.
Human attention is among the scarcest resources in the software factory.
One practical note if you use Conductor: Conductor Cloud went read-only on Nov 13 and shuts down Nov 20, so move to local before Friday.
🔍 Verification tooling is having a moment
The "review is the bottleneck" thesis is showing up as shipped products, not just slide decks.
mabl Introduces Agentic Testing Harness lets teams verify agent-written code with guardrails and see change impact before opening a PR. – mabl
Codenotary Releases AgentMon Live adds a "Done Gate" for Claude Code that checks whether agents actually finished; in a two-day eval across nine projects it flagged potentially unfinished work 59 times, and the agent did more work in 12 of those cases.
NeoSigma's Neo is now a plugin for Claude Code and Codex, connecting production traces and failure analysis directly to the code that caused the failure.
Google open-sourced the Elephant-Goldfish Model plugin, which enforces spec-first coding and catches wrong assumptions "in a Markdown file, not in a pull request."
NTT is coming at the same problem from the legacy side. NTT transforms existing software assets into AI-ready development assets reports its traceability tech lifted inconsistency detection from 27% to 80% and surfaced three previously unknown latent bugs. Making old code legible to agents is unglamorous, and probably where a lot of enterprise value actually lives.
🧰 Tooling churn worth a scan
GitHub Copilot moved into Xcode and JetBrains Air: Copilot hit Xcode's built-in assistant in public preview (Nov 4) and JetBrains Air via the Agent Client Protocol (Nov 3), turning Copilot into a portable agent layer rather than a VS Code extension.
JetBrains gives every IDE a workspace for multi-vendor agents: the 2026.3 releases add Air across PyCharm, GoLand and the rest, supporting Claude Agent, Codex, Gemini, Copilot, Junie and OpenCode with your existing keys.
VS Code 1.143 makes the Copilot harness the default and adds Compare Models with an optional Judge model to score attempts.
Copilot's AI approver, a GLM-5.3 leak, Codex on Vertex: Cloudflare disclosed that 155 Workers AI responses from GLM-5.3 (Aug 27 to Sep 9) contained other customers' data due to an SGLang serving-cache issue, with no evidence of deliberate exploitation. A sober reminder for anyone leaning on hosted inference.
UC Berkeley and Microsoft Research's TokenEconomizer cut average task cost by 18–31% across 277 SWE-bench tasks by reusing project knowledge and pruning unused tools.
Two more for your radar: Cursor and Google gave coding agents a voice with two-way spoken back-and-forth while the agent works, and Recursive launched Flint, a free, offline laptop coding model the company says was built largely with AI help.
That's the issue. The conferences and the product churn tell the same story from two directions: generation is cheap, judgment is the moat. Build your review loop accordingly.

