Change your mind mid-build.
Real requirements move. When you redirect a build in progress, agents replan around the change - they do not throw away everything and start over.
A redirection touches only what it needs to.
Each instruction is diffed against the current build plan. Agents recompute the affected tasks and leave the rest of the system in place.
"Make auth role-based."
The schema agent adds a role to memberships, the backend agent regenerates only the affected route guards, and the frontend gains a role check where it matters. Untouched work stays as it was.
"Switch the database to Postgres."
The infra agent swaps the provisioned database and connection config, the schema is retargeted to the new dialect, and the backend re-validates its queries. The frontend is unaffected.
"Add an activity log to tasks."
A new model is appended, one set of endpoints is generated, and a panel is added to the task view - built on top of what already exists rather than regenerating it.
A replan is a diff on the task graph.
The planner compares the new instruction to the existing plan, marks the tasks that changed, and reruns only those - keeping completed, unaffected work intact.
> make auth role-based[plan] diffing against current graph[plan] affected: schema, backend[plan] unaffected: frontend, infra[schema] + Membership.role (enum)[backend] regenerating 3 route guards[backend] 9 endpoints kept as-is[done] replan applied - no full rebuild
A change should cost only what it changes.
Completed work is preserved
Tasks that are already done and unaffected by your change are left exactly as they were - no needless regeneration.
Only the diff is rebuilt
The planner recomputes just the tasks the change touches, so a redirection is fast instead of a full restart.
The system stays consistent
Because agents rebuild against each other's current output, a mid-build change never leaves layers out of sync.
Build it, then steer it.
Start a build and redirect it as you go - the agents will replan around you.