The Feature Planner writes its own backlog, behind a human gate

The factory built in public at software-factory.dev (opens in a new tab) runs an automation called the Feature Planner, and the part I keep coming back to is what it does when the backlog is almost empty.

It writes its own next tickets.

Triage, every hour

Most of the time the Planner does the boring, necessary job. It triages issues. Its trigger is a cron line, 0 * * * *, so it wakes every hour, plus a manual run on demand.

Each run, it reads the product spec, the architecture doc and the conventions doc, then finds open issues with no status label. For each one it judges whether the issue has enough detail to build. If yes, it normalizes the title and labels it for the queue. If it is too thin, it adds a needs-human label and posts a comment asking specific questions.

Beyond-Spec mode opens under three conditions

Here is the detail worth copying. The Planner has a mode called Beyond-Spec, and it only opens when three conditions hold at once:

• No unimplemented spec items remain.
• The backlog has fewer than three open items.
• Fewer than five issues are already parked on needs-human.

Only then does the factory start proposing new enhancement work itself. And even then it splits the proposals. Low-risk polish goes straight to the backlog. Anything high-risk, a new feature or a schema change, gets needs-human and waits for a person to comment “approved”.

The gate is the whole thing

An automation that can invent its own work is useful right up until it invents the wrong work. The fix is not to forbid ambition. It is to let the factory propose freely and route the risky proposals through a human first.

The pattern travels. An automation that only reacts to a human-written backlog stalls when the humans get busy. One that writes its own with no gate hands you ten speculative tickets by morning. The condition that opens Beyond-Spec mode sits between those two.

The automation itself: feature-planner.yaml in the factory’s repository (opens in a new tab).

When your factory runs out of queued work, who decides what it builds next, and what stops that decision from going too far?

Further reading

• Human accountability →
• Agentic Software Factory: Advanced Team Training →