Agents build. People decide.
About 50 agent sessions run at once across three production systems. This is how that stays safe and fast.
How a change ships
Seven steps. Three of them only a person can do.
- A person
A person writes the spec
Drafted with Claude inside the app, then reviewed section by section.
- A person
Only a person can accept it
Agents can file work in the queue, but only a person can accept it.
- An agent
An agent builds it
Claude Code or Codex picks up the item and works in its own copy of the code and database.
- An agent
A preview link that opens on a phone
Every change runs at its own private link with its own copy of the database. Email and payments are switched off outside production.
- An agent
The checks run
Automated tests and checks run on every change, 209 steps on Brain alone. Past incidents have checks of their own.
- A person
A person merges
Agents can’t merge their own work. A person reads each change and decides.
- An agent
Merging is deploying
The new version goes live on its own, linked back to the spec and the session that built it.
The rules
Agents draft, people send
Anything that moves money, emails a customer, deletes data or widens access waits for a person. Everything else an agent can do on its own.
Keep agents busy, not waiting
Whatever needs a decision comes to the top of one queue. Everything else keeps working in the meantime.
No lock-in to one AI company
Claude Code, Codex and ChatGPT work from the same queue, and any AI feature can change model without a release.
Every incident leaves a check behind
When something breaks, the fix ships with an automated check that stops it happening again.
Works from a phone
Every system has a Claude connector and every change has its own preview link, so all of it can be reviewed from anywhere.
Patterns running in production
The build fails if Claude can’t use a feature
Every screen and API needs an MCP tool, or a written reason why not, before it can ship. That is why the tool counts are so high, and why an agent can do almost anything a signed-in person can.
501 MCP tools across Acquirelytics and Brain
Errors file themselves as work for agents
When something breaks in production it becomes one written-up ticket, ready to hand to an agent. An hourly sweep sorts the new ones, and no agent starts a fix until a person accepts it.
Any AI feature can change model without a release
Each AI feature has its own model and prompt, and changing either is a setting that applies on the next call. Every call is logged with its cost and speed. At Acquirelytics a new model is tested against people’s own decisions before it takes over.
Many agents in one codebase without collisions
Each agent works in its own copy of the code, with its own database and preview link, and can’t commit into another agent’s copy. The house rules are thirteen practices, each written after the incident that taught it.
~15 ms to give a new agent its own copy of the database
The agents have their own GitHub account
Agents commit as their own GitHub user and go through the same steps as everyone else: the automated checks, then a person reading and merging the change.
500 pull requests opened by the agents’ account, each merged by a person
Starting an agent takes a hardware key
Starting or steering an agent session needs a tap on a security key or passkey, so a compromised server can’t set an agent going on its own.
Get in touch
Building something similar, or want to work with us? Email [email protected]. For consulting in the UK, see FireClaw.