8 Days, 3 Agent Teams: What Changed and What It Costs

Between 10 and 17 September 2026, three of the biggest names in AI coding shipped what is recognisably the same product. Not a coincidence of marketing — the same architecture, within eight days.

If you use any one of them, the other two tell you where it is going. Here is what they actually share, and what that means for the person paying.

1. The coordinator does not write code

This is the real change, and it is easy to miss because it sounds like a detail. In all three, the top-level agent never touches the codebase. It reads the goal, decides what the jobs are, hands them out, and reads the results back.

• Cursor says it plainly: the coordinator "doesn't write code itself but directs other agents that do" — and because it only delegates, it is never blocked and stays responsive while work runs.
• Claude Code splits a goal into threads, each a separate cloud session working its own branch and its own copy of the repository.
• Devin shards the codebase into focused batches, runs parallel agents over them, then has a final agent merge, de-duplicate and rank the findings.

It also explains why they all arrived together. A single agent is limited by its context window — the longer a job runs, the more it forgets. Splitting the job across many agents, each with a small focused context, is the obvious way around that ceiling. Devin says so directly: the point of sharding is to keep "each agent's context focused" and spend the reasoning budget on relevant code instead of repeated searching.

2. Everything runs in the cloud — that is the catch

All three execute on the vendor's machines by default. That buys you real things: work continues when you close your laptop, and you can run far more agents in parallel than your own machine would allow.

Cursor is the most flexible of the three: cloud by default, but it will spin up a local agent when something has to be tested on your machine. Worth knowing if you work on anything that cannot be uploaded.

3. The bill grows while you are not watching

This is the part to sit with before you turn any of it on, and it is where the money angle is genuinely different from a normal coding assistant.

A chat assistant costs you tokens while you are typing. A coordinator with parallel workers costs you tokens per worker, in parallel, for as long as the job runs — and Cursor's Projects can be told to work without being prompted at all: watch a Slack channel, run on a schedule, follow every pull request and fix CI when it breaks.

• Start on a small, bounded job — one migration, one scan — and read the token usage afterwards before you trust it with more.
• Pick the model per job, not per tool. A coordinator reasoning about what to do needs a good model; workers doing mechanical edits often do not. Check whether your tool lets you set them separately.
• Turn scheduled or event-driven runs on last, once you know what one manual run costs. A coordinator that wakes on every PR is the fastest way to a surprise.
• Watch the cache line. These systems re-read the same repository constantly, so cached input is most of the bill — which is exactly why Anthropic cut Fable 5.1's cache reads to $0.25 per million.

Which one, if you have to pick

Honestly: whichever is attached to the tool you already use. The architectures are close enough that switching editors to get one of these is not a good trade in September 2026. What differs is fit.