The Coordinator Wars: What September Actually Changed in Agentic Coding

Quiet month for model releases, loud month for everything around them. Opus 5.5 and Sonnet 5.5 turned up, both solid, neither surprising. Same context window, same pricing tier, a benchmark bump that Anthropic measured on its own scaffold so I’ll believe it once someone outside the building confirms it. Fine. Switch the defaults, move on.

The thing that actually kept me up reading changelogs was Claude Code Projects. Anthropic has rebuilt it around a coordinator directing parallel agent threads, each one a full cloud session on its own branch, sharing memory and a common file library. You describe the work, the coordinator splits it up, the threads go off and do it, including opening pull requests and babysitting CI until it’s green. That’s not a chat window with extra steps. That’s a small team of juniors who don’t sleep.

I’m intrigued and a bit wary in roughly equal measure. Intrigued because the merge-conflict handling sounds sane, they’re treating concurrent threads the way you’d treat concurrent PRs, not pretending the problem doesn’t exist. Wary because “running multiple threads can cause Projects to reach plan usage limits faster” is doing a lot of quiet work in that sentence. Translation: this will eat your subscription if you’re not watching it. I’ll pilot it on something scoped and parallelisable, a repo migration maybe, and keep one eye on the usage meter the whole time.

What’s interesting is Cursor is building the exact same thing from the other direction. Rollouts, Security Reviewer, coordinator agents that watch a Slack channel or a PR queue and delegate work as it comes in. Different company, identical bet: the war isn’t about who has the smartest single model anymore, it’s about who owns the layer that manages a fleet of lesser models doing the grunt work. Genuinely fascinating to watch two vendors converge on the same architecture independently within weeks of each other. Also a little exhausting, because it means I need an opinion on orchestration now, not just on which model writes better Python.

The unglamorous but more important release, if I’m honest, is Gemini CLI v0.60. No new features, just a security patch batch: a hardcoded API key removed, RFC 9207 issuer checks added to the MCP OAuth flow, tighter path and symlink validation, provenance checks on untrusted tool output. If you’re running Gemini CLI against third-party MCP servers or repos you don’t fully trust, this is the kind of release you apply the day it lands, same as you would with OpenSSL. Nobody writes excited blog posts about this stuff and that’s exactly why it matters.

One thing I keep turning over: all this agent-fleet infrastructure is impressive engineering, and every one of those parallel cloud threads is burning compute somewhere. Nobody in these release notes mentions the power bill. I don’t have a tidy conclusion there, just a nagging feeling that we’re scaling up usage faster than we’re scaling up any conversation about the cost of it.

Google’s Antigravity, for what it’s worth, is still digging out from its own hole after the May update that replaced people’s IDEs without asking. Plan review and a plugin marketplace this month are the right moves, but trust doesn’t come back on a changelog’s schedule. I’ll give it another quarter before I recommend it to anyone who got burned.

Net effect of the month: less “which model is smartest” and more “who’s managing the agents that manage the agents.” I don’t know yet whether that’s progress or just complexity wearing a nicer jacket. Probably both.