See how each agent read the problem, what it built, how a blind critic judged it and why, and what is running now.
Agentic engineering is how a lot of software gets built now: agents write the code and people direct them. Lucid Coding is the practice of doing that with your eyes open. You know your system, your agents' thinking, and how each piece was built and judged.
Coordinator is built for that. It does not pick your stack or your screens. It brings the process: a bar agreed before the build, a critic who never grades its own work, hard stops in code, and a record of every step you can open in a browser.
Get the work done first. Speed and cost improve with time and experience, and each run's record is where that experience comes from.
Every run follows the same six steps. Each one leaves an event in the run's log, so the dashboard can show it and a crash cannot lose it.
The lead reads the project and the task, researches, staffs builders and critics from your harness's own agents, and asks every open question in one batch, its recommendation first. Then it runs without you.
Each unit of work gets a bar a critic can check: behaviours written as given, when, then, a failing test that must pass, or a reference to match. Bar changes stay on the record.
One fresh builder per unit, each in its own git worktree, in parallel where units are independent. Mechanical work can go to a cheaper model; critics and work on security, money, migrations and auth stay on the strong one. Each round records which model built it.
A separate critic, who never sees the builder's reasoning, checks the real result against the bar and marks each behaviour with evidence. A detector reads the diff for skipped tests, removed assertions and rewritten snapshots.
run.mjs next decides what happens to each unit: build again, change strategy, or stop. Round and time budgets, no-progress and same-cause detection are enforced by the script, and steps taken out of order are refused.
The retro turns repeated gaps into lessons quoted into the next run's briefs, tunes the budgets from past runs, and proposes changes to the skill for a person to review.
Real screenshots from real runs. Coordinator builds itself with Coordinator, so most of them are its own release runs.
A question comes once, with the lead's pick marked Recommended. Answer it here, from a phone, or in an issue comment. A unit waiting on another one says Waiting for A and links to it; only a real question for you turns amber and says Needs you.
The bar, as behaviours a critic can check one by one. Screenshots, test output and logs attach to the behaviour they prove, and run.mjs adopt copies them into the run before a worktree is deleted.
Every round keeps the builder's report, the critic's gap, its root cause and a score. Here a critic failed the model-routing unit three times, on three different gaps, and the script stopped it when its round budget ran out.
Each map is drawn from the run's events, with no extra input. Every box carries a status mark, its id and title, so nothing depends on colour alone.
run.mjs serve serves it on 127.0.0.1 behind a token, updating in place as the run moves. A page per unit, timelines with search and filters, and Export PDF on every page./coordinator answer q3 b in an issue or PR comment. Only people with push access can answer.reins_perf), inspect DOM and CSS layout (reins_dom) and snap panels by id instead of clicking by pixels. A reins_report in Markdown, attached to the unit, shows the critic what was slow, duplicated or clipped, front end and back end, and a before/after compare proves a speed-up with numbers.run.mjs quality: duplication, long functions, nesting, naming, security and PDPL/GDPR privacy signals, each with the file and line to review. Signals for a person, not a certification.--continues run:unit links them, and the stats count the chain as one unit./coordinator:init learns the project once and asks where state goes, the theme, budgets, the merge rule and the hosted app. /coordinator:doctor checks the setup, one line per item with a fix.AGENTS.md block; with no question tool there, questions go to the dashboard.helper.mjs setup. Unsure answers escalate to the strong model; a helper never judges.
run.mjs quality, each one for a person to check.
After a crash, a closed session or a week away, the run is still there. Every step is an append-only event log in .coordinator/runs/<run>/, with the questions you answered, the bar, every round and its evidence.
Work in flight is checked against its branch or output; a unit with nothing to show goes back to build. Start a new run for new work, and use --continues when it finishes a unit an earlier run left open, so that unit's history comes along.
/coordinator:doctorchecks the setup first: Node, gh, settings, stale locks and dashboards/coordinator:statuswhere each run stands; the lead carries on with run.mjs next and does not ask your answers againrun.mjs pull <run>on another machine: restores the run from the coordinator-state branch the lead pushed at its last pit stop; a public repo keeps it in a private stateRepo instead/plugin marketplace add mindmakr/coordinator
/plugin install coordinator@coordinatorThen, in a project: /coordinator:init, then /coordinator:run Fix issue #142. Needs Node 18+ and git, nothing else. Add .coordinator/ to your .gitignore.
In OpenAI Codex CLI: copy the skill folders into .agents/skills/ and add the AGENTS.md block. Install guide
Follow runs on the web, answer from a phone, and get alerts. One container and Postgres, with your own GitHub App.
Self-host guideThe same service, run by us, free during the beta. What it reads and keeps: permissions and privacy.
Follow the hosted betaGive your agents hands and eyes inside your app: typed actions, state, DOM/CSS diagnostics and panel screenshots, and one timeline from the click through the server and database back to the screen. Session reports in Markdown become a critic's evidence; testers can export theirs for an agent to analyse. Stripped from production builds.
Explore reins