Letting coding agents merge to production

FH//AGT

two to five days on your repo · €1,000 a day · team enablement

Far Horizons helps engineering teams let AI coding agents, like Claude Code and Codex, merge their own work to production without it going wrong. It's two to five days on your repository with your engineers, setting up the build, hooks and merge rules that catch what an agent gets wrong before a person has to.

Make the build the gate

day one · on your CI

Anything you want enforced on an agent is enforced by failing the build, not in PR review, because nobody reads all of the code any more. More tests, end-to-end checks of what the user actually sees, and an AI reviewer whose comments must all be resolved before merge. The build can take an hour: the agent doesn't mind waiting, and goes round again when a check fails.

Hooks that redirect

days one to two · in your agent setup

When an agent does something dangerous, like writing raw SQL against a database or editing a committed migration, a hook stops the command and tells it what to do instead. A flat block makes the agent improvise around it. Every repeated mistake becomes a hook or a lint check, and the hooks are reviewed whenever you change models.

Shipped means deployed

day two or three · merge queue and deploys

A merge queue so parallel agents don't thrash the pipeline, and one rule: the work isn't done until the deploy on the merge commit is green, acceptance checks included. Production sits behind a production-like environment, and new features go out switched off behind toggles.

Decide what merges without a human

last day · with your team

Not everything should auto-merge. The AI scores each PR for risk, low-risk changes merge on green, and people review the ones that need them. Database access stays read-only. Your engineers leave with the setup in the repo and the habits to keep it up to date.

Learned the hard way

Far Horizons' own product is where the agent gets more access and freedom than it would on client work, on purpose, to find where it breaks. It has dropped a production database (the backups made it a non-event) and run commands that skip migrations. Each time the fix was in the environment, not a sterner prompt. That's what gets set up on your repo.

Who leads it

Far Horizons OÜ is the AI engineering practice of Luke Chadwick, based in Berlin. He has shipped software since 2001, is a continuous-delivery person from before pull requests, and writes code with agents every day. There's more on the about page.

Before you write

Is it safe to let AI coding agents merge to production?

Yes, if the build is strict enough to catch what the agent gets wrong, and if you treat each mistake as a gap in the build rather than in the prompt. The agent has no sense of consequences, so the safeguards have to be in the environment. Far Horizons runs its own half-million-line product this way: agents merge their own PRs through fourteen required checks, and one overnight batch merged twenty PRs, ending with the last deploy green on every check.

What does it cost?

This is team enablement: two to five days working on your repo with your engineers, at €1,000 a day. It is scoped and priced before anything starts.

Which agents does it cover?

Claude Code and Codex mostly, and the same approach works for others. The models behave differently, so the guardrails are tuned to the ones you run, and they are reviewed when you switch.

Our codebase is old and the test coverage is thin. Can we still do this?

You can use agents on it, but they'll need more hand-holding, and auto-merge has to wait until the tests can tell when work is done. If you're not sure where you stand, the two-week codebase audit answers whether it's safe to point agents at it.

Who does the work?

Luke Chadwick, who has run coding agents merging to production on his own product since January 2026 and has built production LLM systems since 2022. He works from Berlin with teams across Europe.