Claude Code finds its footing as agent platforms grow up

Agents that plan, delegate, and stay on call

The week's biggest swing came from Cursor's new Projects surface, which puts a coordinator agent in the left-hand nav and lets it plan a body of work, farm it out to thousands of parallel subagents on cloud machines, and hand back finished results for review. Shared context files sync across cloud and local machines, and subscriptions let the coordinator respond to Slack channels, schedules, or pull requests without a human prompt. It's a clear bet that the unit of software work is shifting from "a session" to "a standing project."

Cognition pushed Devin in the same direction, less from ambition than from operational reality. The September release adds plugins that bundle skills, rules, hooks, and MCP servers into one installable package with security notices on every install, and archiving a session now closes the open PRs of child sessions archived with it — exactly the kind of hygiene you need once agents spawn agents. A few days later, PagerDuty oncall integration and 21 zero-config OAuth MCP servers arrived, letting Devin act as an oncall responder that triages incidents and posts investigation notes, plus an idempotency_key on the v1 Sessions API that returns the original session on repeat calls.

A hard week of hardening in Claude Code

Claude Code's release cadence dominated the week, and the story was less about features than about surviving its own velocity. A gateway sign-in regression capped off a bad run: the undocumented CLAUDE_CODE_USE_GATEWAY variable began forcing Cloud-gateway sign-in on its own, breaking every request for proxy and LLM-gateway setups using an API key or custom auth headers. Behind it came another fix for HTTP 400 failures on third-party Anthropic-compatible endpoints caused by a regex in the Artifact tool's input schema, plus a WebFetch hang and a busy loop pinning a CPU core.

The recovery releases carried real substance too. v2.1.265 added --plugin-dir support for a whole folder of plugins and a 1 GB cap on tool results saved to disk, alongside fixes for prompt-cache reuse in resumed subagents. v2.1.267 added a maxEffortLevel cap across every provider and patched a sandbox containment bypass for marketplace paths with backslashes. And v2.1.269 introduced claude plugin eval for scored, reproducible plugin eval suites and /output-style [name] switching over Remote Control and headless sessions — the most substantial capability drop of the seven, though it needed a follow-up to stop read-only git commands from prompting for permission mid-session.

Codex widens its surface

OpenAI spent the week extending Codex rather than rebuilding it. Codex 0.154.0 brought GPT-6-Astra to the model picker and Bedrock catalogs and added experimental worktrees for isolated checkouts via --worktree. The Python SDK's 0.154.0 added ExternalMessage for sync and async runs, new max and ultra reasoning-effort values, and a breaking change wrapping hook handlers in .root. On mobile, ChatGPT for iOS gained @ mentions and background worktree setup with a Live Activity, while the desktop app brought Quick Chat to the floating Pets controls and Appshots to Windows.

Smaller but useful: Grok Build fixed headless prompt hangs and wrong-session resumption, then tidied its session header and tmux pane lag.