Agents already write our code.

Circus builds customizable agentic workflows, and its core stays independent of whatever consumes it. An ingress layer classifies every incoming Slack message, GitHub event, and Jira update. A control plane ranks and admits the work. A warm pool of running pods means no task waits on a cold start. A runtime hands each task its own isolated environment. An egress layer posts the result back wherever the conversation began.

It is deliberately model-agnostic, driving Claude Code, Codex, and opencode behind one interface and evaluating all of them against the same suite of real historical pull requests. Models change underneath us constantly. The platform is what compounds.

---
cron_spec: "23 14 * * 1"
disabled: false
owner: "<engineer>"
memory_folder: "background/<job>"
---

# then the runbook, in plain English

Every scheduled agent is a file like this, with a named owner. Creating one is a pull request. Changing what an agent believes is a diff anyone can read.

The product still gets built, faster than ever. But a fifth of everything our engineers ship goes into the machinery that instructs, reviews, and constrains the agents — and that is the part we expect to grow.

The leverage is why people stay. A fix here is rarely a patch; it is a change to what every agent knows, inherited by every run that follows. A good afternoon does not fix one bug. It retires a category of bug permanently.

What that work is

Design an agent

Decide what it owns, what it needs to see, and how you will know it worked.

Build the runtime

Extend Circus itself: ingress, scheduling, isolation, tracing, tool access.

Evaluate

Build the eval sets that decide which model we trust for which work.

Encode judgement

Turn what a good engineer knows into something an agent applies every time.

Our engineers touch dozens of pull requests a week themselves, while the agents they spawn merge many more. The code you write moves up a level — policy engines, eval harnesses, diff analysis, schedulers, admission control — and there is more of it, not less.

The environment bootstraps itself, demo data included. Every pull request ships with a live instance of the platform running that branch, so reviewing means clicking a link — and you can take over a running agent session whenever you want the wheel.

The conventions for whatever you touch sit in the folder you are touching, compiled from everything the system has already learned. That holds for you exactly as it holds for an agent, and it is why people get productive here in days.

Onboarding itself is run by an agent. When your question exposes a gap in the documentation, closing that gap is part of the answer — so nobody hits the same wall twice.