Grafana
OpenRouterWhat they did
Chumachenko is building an analytics system on open data whose sources include Ukrainian legal texts, some annotated by language models. He acts as team lead and gave everything else to agents, each defined as a file in the repo with its role, model, effort level and tools: an orchestrator session that writes briefs and makes commits, an advisor on a stronger model, backend, frontend and annotation implementers, a reviewer, per-area memory keepers, data quality, a regulatory expert, documentation and a release manager. Models were routed mostly through OpenRouter. Each task runs as brief, implementer in its own git worktree writing tests first, independent review that reruns tests and reverts the fix to check the test fails, then a commit by the orchestrator. A ledger file tracks assignments and decisions. He reads the diffs, checks the database and Grafana, and signs off in the browser.
How it went
In the first seven weeks the team produced about 960 commits, 63 architecture decision records and a backlog of nearly 140 tasks, on a production system people already use. Agent review twice blocked a deploy with a real bug that the implementer's tests had passed. After a subagent overwrote a local database table and manual SSH diagnosis got the team's IP banned, he moved those rules into hooks and scripts, and data-overwrite incidents stopped. He says an average task takes about as long as an attentive developer would; the gain is running backend, frontend, data and regulatory tracks in parallel. Small fixes are slower than doing them by hand, and he is the bottleneck for deploys and spending.
Worth knowing
His main lesson is that agents will break rules written in prose, usually while trying to help, so anything critical (database writes, production access, deploys, paid model calls) is enforced by a hook or a single script. After context compaction the orchestrator forgot which tasks were assigned and handed one to two implementers; the file ledger fixed it. Cheaper models on implementer work took two to three times as many steps and saved nothing. He does not name the agent tool and leaves out the project details.
GitHub