AI Catchup

The AI Catchup -- May 12, 2026

By 4 min read

Welcome back. The throughline this week: agent runtimes are getting headless control surfaces, the dev-tool side is paying down quality debt, and one major vendor API has a deadline you do not want to miss.

Let us get into it.

This Week in AI

The big story is headless agent control as a first-class entrypoint. Codex CLI 0.130.0 shipped on May 9 with a new top-level codex remote-control command that starts the app-server in a headless, remotely controllable mode. It is a simpler entrypoint than launching the app-server indirectly, and it is aimed at teams that run Codex as a service or integrate it into other tooling. The same release adds client-side paging for large threads (unloaded, summary, or full turn item views), bundled-hooks visibility in plugin details, link metadata and discoverability controls for plugin sharing, and Bedrock auth that picks up AWS console-login credentials from aws login profiles.

That follows Codex CLI 0.129.0 two days earlier on May 7, which brought modal Vim editing to the TUI composer via /vim, a redesigned resume/fork picker with raw scrollback and workspace-aware /diff, a new /hooks browser with before/after compaction support, workspace sharing and access controls for plugins, theme-aware status lines, and Codex Apps auth surfaced through Guardian. Together the two releases tell a clear story: the TUI is getting tighter for hands-on use, and the app-server is getting more controllable for everything else.

Tool Spotlight

Claude Code 2.1.133 shipped on May 7 and is the kind of release that does not get a marketing push but matters every day. The headline change is a new worktree.baseRef setting (fresh | head) that defaults to fresh, which moves EnterWorktree's base back to origin/<default> after several months of branching from local HEAD. If you have been hitting unexpected merge conflicts on agent worktrees, this is why -- and the default is now back where it should be.

The same release routes HTTP(S)_PROXY, NO_PROXY, and mTLS through the entire MCP OAuth flow -- discovery, dynamic client registration, token exchange, and refresh -- not just the initial request. It exposes effort level to hooks via $CLAUDE_EFFORT, adds Linux sandbox path overrides, and fixes a refresh-token race that was 401-ing parallel sessions. For enterprise teams behind a proxy with MCP servers behind their own auth, this is the release that finally makes the whole flow work end-to-end.

What We Are Watching

Google is rolling out breaking changes to the Gemini v1beta Interactions API with two hard deadlines you should put on the calendar. The new schema replaces the outputs array with a steps array, removes response_mime_type in favor of a polymorphic response_format, and introduces new streaming event types. For REST users, the new schema becomes the default on May 20, 2026, and legacy behavior is removed entirely on June 8, 2026. Older Python and JS SDKs (1.x) also break on June 8. If you have not audited which version of the Gemini SDK is pinned in production, this week is the time.

Comparison Reading

The Codex CLI vs Claude Code vs Cursor architecture deep-dive remains the right anchor for understanding why this week's releases shape up the way they do. Codex's app-server design is what makes remote-control a clean addition. Claude Code's MCP-centric model is why proxy and mTLS coverage across the OAuth flow is a release-worthy fix on its own. The architectural choices each tool made a year ago are still driving what each one ships today.

Quick Hits

  • Cursor enterprise controls (May 4): Cursor added granular model and provider access controls, soft spend limits with automated alerts, and richer usage analytics that break consumption down by product surface (Cloud Agents, Bugbot, Security Review). Existing blocklist customers need to migrate by June 1.
  • Codex CLI 0.130 plugin metadata: plugin details now show bundled hooks, and plugin sharing exposes link metadata plus discoverability controls -- a small change with real implications for plugin trust as the ecosystem grows.
  • Codex CLI 0.129 /hooks browser: the new hooks browser supports before/after compaction, which is a meaningful primitive for keeping long sessions on-track without losing context state.

That is it for this week. The AI Catchup publishes every Tuesday. If you found this useful, subscribe to get it in your inbox.

Until next week -- stay caught up.

Get the weekly AI Catchup

Tools, practices, and what matters -- in your inbox every Monday.