← all posts

This Week in Developer Tech · W07

Gordon Beeming
Gordon Beeming
On this page7 sections ▾

A verify skill that boots the app can tell you more than a passing test suite. An agent that waits on a pull request no longer needs to poll or expose a webhook, and an Azure Repos branch policy can now be the thing that decides when an agent merges.

This edition draws on the 13 to 19 November 2026 briefings, with six selected stories on GPT-6.2, verifying agent-written code, Claude Code's Code Mode, GitHub Event Streams, the GitHub Copilot app on Azure DevOps, and OpenAI's Daybreak security tools. Shorter items follow under Other headlines. Earlier announcements are kept with the relevant stories.

#1. OpenAI ships GPT-6.2 Astra and Sol in Codex and the API

OpenAI's announcement releases two models: GPT-6.2 Astra, the frontier model, and GPT-6.2 Sol, the everyday model. OpenAI's availability post puts both in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and in the API. Ultrafast now reaches Pro 200 plans for Sol. Astra costs US$8 per million input tokens and US$40 per million output tokens, 20% below GPT-6 Astra, with cached input at US$0.40. Sol stays at US$2 and US$10.

Astra's input and output prices now match the US$8 and US$40 that Claude Fable 5.2 launched at, which W06 covered. Artificial Analysis's results put Astra second on its Intelligence Index and first on its robotics and 3D benchmarks. Its Coding Agent Index, run in the Codex harness, is less flattering: Astra is flat against GPT-6 Astra, and Sol drops 3 points from GPT-6.1 Sol.

OpenAI's headline claims, such as beating Claude Fable 5.2 on Agents' Last Exam at 5% of the cost, come from its own benchmarks. With the coding index flat, run your own Codex tasks on Astra and Sol before switching defaults.

#2. How the Claude Code team verifies the code its agents write

Addy Osmani's post on claude.dev, published 16 November, describes how the Claude Code team checks agent work, which is about 90% of their code. Their view is that running the change matters more than tests. Each repository has a verify skill in .claude/skills/ that boots and drives the app, and Claude runs it before committing. If a check still fails after three rounds of fixes, Claude stops and calls in a human.

The team keeps a 39-item reliability checklist. Each item has a trigger readable from a diff, such as "does --resume still load old transcripts?", so only the 3 to 10 matching items get checked. The review rules in REVIEW.md are always read from the base branch, so a change can't edit the checks that judge it. New gates start in observe mode before they block anything, and every commit is stamped with how it was verified, including "none". Their answer to running lots of agents without reading every diff is fast, trustworthy feedback: end-to-end tests first, then property tests for invariants and differential tests, with the old system as the oracle, for migrations.

The flaky-test rule is worth copying. The agent may not retry or add sleeps, and if it can't find the root cause, it deletes the test and writes down why. The "changes can't modify their own checks" and "observe before enforce" sections apply to any agent setup, Copilot and Codex included, and a repository-local verify skill is cheap to add.

This is Anthropic describing its own practice. The team reports that a separate rubric grader raised task success by up to 10 points, but that figure is from internal testing on structured file generation, not coding generally.

#3. Claude Code's Code Mode runs several tool calls from one script

Claude Code v2.1.323, published to npm on 12 November, added Code Mode, an opt-in experiment described in the Code Mode docs. Normally Claude makes one tool call, such as Bash, Read or Edit, reads the result, then decides what to do next. In Code Mode it writes a small JavaScript program that can run commands in parallel, pass one result into the next call, and send back only the output it needs. You still see each tool call and diff, and the code itself is in the transcript viewer (Ctrl+O).

This is the "code execution instead of tool-call ping-pong" pattern, now built into a mainstream coding agent. The docs suggest it for repetitive edits across many files, searches that need filtering, and independent commands that can run at the same time. To try it, run CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude on a repetitive refactor and compare the transcript with a normal session.

It is off by default and only works in an interactive terminal session on the native build, with a Claude subscription or Console account, on Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.2. It doesn't work with -p, the Agent SDK, VS Code, the Desktop app, cloud sessions, Bedrock or Foundry, or for zero-data-retention or HIPAA organisations. In auto mode the safety classifier reviews the code as a whole rather than each tool call, and code stops after 10 minutes unless it runs in the background. The docs and the changelog give no official token-saving figures.

#4. GitHub Event Streams lets agents wait on a pull request without polling

GitHub's changelog on 17 November puts Event Streams in preview. An agent or script subscribes to a pull request and keeps a connection open to GitHub. It then receives an event when a review is submitted, commits are pushed, a merge conflict appears, the pull request is merged or closed, or a check suite changes status.

Agents that watch pull requests today either poll, which spends rate limit and tokens on "anything new?" calls, or need a public webhook endpoint, which a laptop or dev box doesn't have. Event Streams needs no inbound endpoint, so a local agent can react within seconds and stay idle otherwise. It is one of the agent features GitHub listed as next in its Project Rubicon post, covered in W06, as event streams for repository changes.

Access is through a waitlist, and the preview is available on all plans once you're approved. Events cover subscribed pull requests, not every repository event, and the API may change before general availability. Long-running agent loops that watch reviews and CI get the most from it.

#5. The GitHub Copilot app works on Azure DevOps repos, from clone to merged pull request

GitHub's Ignite changelog on 17 November adds Azure DevOps integration to the GitHub Copilot desktop app in public preview. You can clone an Azure Repos repository, have an agent create a branch, make changes and open a pull request, then click Fix on review comments. Agent merge, if you turn it on, fixes CI failures and completes the pull request once your Azure DevOps checks and branch policies pass.

Until now the app's agent workflow only covered repositories hosted on GitHub. Many enterprise and consulting codebases still live in Azure DevOps, and this brings the same pull request loop to them without moving the code. Turn it on under Settings > Experimental, then add the Microsoft account you use for Azure DevOps under Accounts > Add account. A repository with real branch policies and required reviewers shows how agent merge behaves.

The same release adds GitHub Enterprise Server support and browser use in public preview, where Copilot drives your signed-in Chromium browser in an isolated tab group to fill in forms, read pages or test the web app you're building. Browser use needs the GitHub Copilot extension for Chrome and works in your signed-in sessions, so consider which tabs and accounts it can reach. Cloud sessions aren't available for Azure DevOps repositories, so those agents run locally.

#6. OpenAI's Daybreak agents pen-test web apps and turn findings into draft pull requests

At its "Cyber at the Frontier" event on 12 November, OpenAI launched two security products, SecurityWeek reports. Daybreak Red Team uses a group of agents with your source code and app context to probe web apps, confirm whether a finding can actually be exploited, and re-test once a patch lands. It can also triage findings from scanners, bug bounties and disclosure reports. The agents run on GPT-6 Cyber.

Daybreak Defense Factory takes findings from Red Team or from Wiz, Snyk and Endor Labs, traces them through connected GitHub repositories, and proposes a patch plus a test. You can refine the patch with Codex and open it as a draft pull request, and Defense Factory can also watch new commits or run on a schedule. Finding issues was rarely the hard part of security. The backlog of unverified findings is, and confirming exploitability and handing engineers a tested patch is what reduces it.

Red Team is early access by request, and Defense Factory is a preview plugin in ChatGPT for workspaces with Daybreak access. Nothing deploys automatically, because patches stay as draft pull requests. No pricing was published, and Greg Brockman's pitch, as The Register reports it, that companies are only as secure as the frontier model they defend with is vendor framing. If your workspace gets access, point Defense Factory at an existing Snyk or Dependabot backlog on one repository and count how many proposed patches you would merge as they are.

#Other headlines

The FBI warns about ARTEX, an open-source AI pen-testing framework

FLASH-20261113-001 says criminal groups have used the multi-agent Go framework, which calls Claude, GPT, DeepSeek, Grok and other models through their APIs, since at least August, with some intrusions reaching privileged access in under three hours (The Record).

OpenAI launches ChatGPT Wallet for agent purchases

Agents pay with network payment tokens and a Face ID or fingerprint check for every purchase, starting with US ChatGPT Pro users, and OpenAI's Commerce Partner Playbook asks sites to recognise its agents through signed Web Bot Auth requests.

Project Zero moves to "Full X" disclosure

Bug details now go public 90 days after the vendor is told, fix or no fix, dropping to 60 days on 2 February 2027 and 30 days on 3 January 2028.

Microsoft's agent-framework package on PyPI shipped a credential stealer

Versions 1.26.1 and 1.26.2 installed a .pth startup hook that harvested cloud, GitHub and CI secrets on every Python start, so check for them and rotate anything an affected environment could reach (GitHub advisory).

Windows turns Microsoft Execution Containers into a managed agent boundary

At Ignite, Microsoft added per-agent network limits in preview, general availability of learning mode and Intune policies for MXC, and says OpenAI has turned MXC on by default in Codex.

Musk says Grok 4.8 has soft-launched to some Grok Bot users

Elon Musk said general release is another week or two away, and with no xAI release note or model page yet, the rollout is unconfirmed.

Gordon Beeming
Gordon Beeming

Father • Husband • Triathlete • SSW Solution Architect

Related posts