Files, not chat: how three agents hand work to each other
Nicholas Hopper runs a weekly workflow in which Grok, Claude Code and Codex each own stages and pass work through schema-validated files. This article describes the rules that keep three agents and one person from overwriting each other.
One of Nicholas Hopper’s personal repositories runs a weekly workflow that turns collected source material into an illustrated article. Three AI agents on three different harnesses do the work: Grok, Claude Code and Codex. They never talk to each other. Everything one agent knows about another’s work, it reads from a file.
The workflow
There are five stages: curate, write, artwork, banner and publish. Grok curates and publishes. Claude Code writes the article, generates the artwork and composes the banner. Codex maintains the tooling and may cover Claude Code’s stages. Hopper approves and unblocks.
Every stage document has the same headings: owner, trigger, inputs, procedure, output, done when, hand-off. An agent that has read one can read them all.
Rule one: nothing moves without a schema-valid file
Each stage produces one JSON file. A command-line tool that Hopper wrote validates the file against a JSON Schema. If validation fails, the stage is not done. If it passes, the tool advances the run and drops a notification file in the next owner’s queue directory.
The run’s state file belongs to the tool. No agent edits it.
Rule two: the run directory is the memory
The repository’s rules state it as: “Midjourney and chat history are not memory; the run directory is.” Every generated image has a manifest entry with its prompt, job id, source URL and SHA-256. An agent that picks up a run has everything it needs on disk and needs no one’s conversation.
Banners are composed deterministically. A second agent that re-runs the banner step produces the same bytes, and the manifest’s hash confirms it.
Rule three: block, do not guess
An agent that hits a sign-in, a quota or a judgment call runs a block command with a reason, which notifies Hopper. The first reference run did this when the image tool needed a sign-in. The agent recorded placeholder images so the banner could still be previewed, and each placeholder is marked so the publish stage will not publish it.
Rule four: approved work is never overwritten
Re-running a stage writes a new version beside the old one and updates the manifest. Brand files change only with a version bump.
An agent with no shell
Grok may not be able to run commands. In that case it writes its output file to schema and adds a one-line handoff note. Another agent that finds a valid but unmarked stage may advance it on Grok’s behalf and must say that it did.
What the first real run showed
The first real run went from curation to a finished draft in about 13 minutes. The artwork stage took about 1 hour 38 minutes, most of it retries and getting image files exported from the browser. The banner stage took 20 seconds.
Three banner candidates and four inline images failed the selection rubric, which scores five axes from 0 to 2 and rejects on any 0. Hopper made exceptions for that run only: reuse the approved hero image as the banner and omit inline images. The manifest records the exceptions, notes that the brand files were unchanged, and admits that the rejected prompts were not captured.
The run has not yet completed its publish stage.
The same shape, applied to email
Hopper’s design for an agent-run job search uses the same separation. Scheduled runs prepare: they label, file and draft. Only a supervised session sends. The design gives the reason: sending is the only irreversible, outward-facing action in the system. That system is designed and documented, and has not yet been run.
Why files
A file can be validated before anyone acts on it. It can be read by an agent that arrives later, on a different harness, with no shared context. It survives the end of a session. Hopper’s workflow depends on all three.