Developer Spends $9,000 on AI Agents and Ships Crossweft, an Open-Source Plugin

A developer who has been building an app with AI agents since January spent more than $9,000 and more than a billion tokens on it. She ended up releasing Crossweft, an open plugin for Claude Code and Codex. Its job is stated plainly: an agent cannot declare work finished until the interrelated parts of the program are in sync. The author told the story and described how the tool works on Habr.

Where agent-written code breaks

Anyone who has asked agents to edit several files at once will recognize the problem. The app itself worked perfectly. But the project has several protection layers, and they drifted apart: something changed in one place and broke in another. The agent edits a file, the local check passes, and it reports "done." Consistency between configuration, tests, rules, and code is left to the human.

The author tried many ways to bring the layers together. One of them was fairly heavy: a separate application that monitored the "heartbeat" of the protection layers and tried to keep them in the same state. It didn't help. Things only came together with Crossweft, where the check is tied to the moment the agent is about to finish a task.

What we know about Crossweft

The announcement that reached our feed reveals only the general outline. It does not say how exactly the plugin finds links between files, which checks it runs, or what it counts as synchronization. Those details remain in the full article, and we will not retell what we have not seen.

What is confirmed:

• Format: open plugin.
• Platforms: Claude Code and Codex.
• Function: blocks the agent's final "done" until related parts of the program are consistent.
• Development cost: more than $9,000.
• Token usage: more than 1 billion.
• Timeframe: the project in which the problem arose has been running since January.

A simple calculation gives roughly $9 per million tokens on average. This is our estimate from the author's two rounded figures. The actual mix of input, output, and cached tokens is unknown, but the order of magnitude is clear. Iterating with agents on consistency tasks is expensive because every attempt means reading and rewriting several files at once.

The approach also has a broader implication. Today most agent workflows trust the model's own report that it has finished. Crossweft moves the completion criterion out of the text the agent writes and into a check the agent cannot talk its way around.

How to apply this

You don't need the plugin to get the same effect, but you need the discipline either way:

1. Write down the links between your project's layers as an explicit list: access policies, configs, tests, documentation, schemas. The agent can't cross-check what isn't described.

2. Make "done" the result of a check, not a claim by the model. A hook or script that returns an error when things are out of sync is more reliable than any instruction in a prompt.

3. Track tokens and money per iteration from day one. Once the total reaches four figures, changing your approach is already expensive.

4. Before installing Crossweft or a similar plugin, open the repository and check the license and the permissions it requests. A tool that intervenes in how an agent finishes its work operates in close proximity to your code.

For teams already working with Claude Code or Codex, Crossweft is worth trying on a small module with linked layers. If it catches even one desync that the tests missed, the installation has paid for itself.