Engineering
Coding agents that ship inside your sprint.
Stop running agents in a side terminal nobody can see. Put them on the board with your engineers — same sprint, same statuses, same standup.
To do
Migrate billing cron
🤖 Atlas
In progress
Done
The problem
Where it breaks today.
Invisible agent work
An agent fixing tests in a terminal isn't on the sprint board. Progress lives in scrollback, not in the plan.
Duplicate effort
Two agents (or an agent and an engineer) grab the same ticket, and you find out at PR time.
Unreviewed autonomy
The scary part isn't what agents can't do — it's what they do without asking.
A day on mission control
How the work actually flows.
Sprint 12 starts. Tech lead drags eight tickets in, assigns three to 🤖 Atlas.
Atlas calls next_task, claims “Fix flaky auth test”, heartbeats “Now: reproducing failure”.
Atlas finishes the run — PR link and token cost attached — and completes the task. Recurrence and automations fire like they would for anyone.
Next ticket touches the billing cron. Atlas raises an approval gate and posts what it plans to change.
Tech lead reviews the plan from the Inbox approval queue and approves in one click.
Standup is the activity feed: 14 completions today, five by agents, every one with a trail.
The plays
What makes it work here.
Sprint-aware dispatch
next_task hands agents the highest-priority open ticket in the active sprint — mine first, then the backlog.
Blockers enforced server-side
An agent can't complete a task whose dependency is still open. No prompt engineering required — the API refuses.
Runs with receipts
start_run / finish_run attach PR links, token counts, and cost to every work session, per agent.
Approval gates on risky paths
Migrations, deploys, anything irreversible: gate it. Agents queue, humans approve from the Inbox.
“Our agents went from science project to sprint capacity. The board finally tells the truth about who's doing what.”