Progressive planning

Start with the outcome

Tell First Mate what should become true, what evidence will prove it and where authority stops. Do not turn every expected implementation step into an Agent instruction.

Human outcome
  → Task
  → inspect current reality
      → Assignment when work is actionable
      → Scout when a bounded uncertainty needs evidence
      → Captain decision when authority belongs to a person
      → dependency when another outcome must land first
  → verified result
  → next distinct outcome, only when evidence justifies one

This is not planning without structure. The outcome, acceptance evidence, ownership and authority are exact. The internal path can change when the repository, tracker, runtime or delivered evidence contradicts an earlier assumption.

Let evidence expose the next work

First Mate reads the authorities that already own the facts before accepting or routing work. Current evidence can expose a prerequisite, narrow uncertainty, make an intended change unnecessary or identify a separate outcome that deserves its own Task.

That last boundary matters: a new Task represents genuinely distinct intent the organization should preserve. AgentOS does not create a decomposition tree merely to narrate every tool call or implementation step.

Separate uncertainty from accepted work

Conversation can clarify intent without creating durable Fleet work. An unassigned Task is deliberate backlog, not accepted asynchronous execution. Active work begins only when AgentOS creates the first accountable Assignment.

When the next action is not honest yet, the organization preserves the uncertainty:

• a Scout returns evidence about one bounded question;
• a Captain decision records a consequential choice that belongs to a person; or
• a dependency states which other outcome must complete first.

Investigation should reduce uncertainty, not create another ship worker racing the accountable owner.

Preserve one accountable owner

Every accepted outcome has exactly one active Assignment owner. If responsibility changes, the Task keeps its identity while an atomic handoff ends the prior Assignment and creates its replacement. History remains intact instead of being rewritten around the new plan.

Direct steering can correct a stale assumption in the live native session. The responsible Mate then reconciles the correction into durable state so later supervision still sees truthful ownership and boundaries.

Keep every fact in its authority

AgentOS connects authorities without copying all of them into a planning database:

• the selected tracker owns human planning and provider workflow;
• PostgreSQL owns durable Tasks, Assignments and Captain decisions;
• Kubernetes owns live workload state;
• Herdr owns native harness sessions;
• an Agent's PVC owns its home and unfinished work; and
• Git and its remote own delivered source and review artifacts.

Raw reasoning, conversation, issue bodies and terminal output stay where they belong. Stable references form the chain of custody. See Native authorities for the complete boundary.

Know what AgentOS does not claim

Progressive planning does not mean that an Agent predicts every future task. It does not replace the project tracker, turn memory into authority or make status activity count as delivery. It also does not remove human judgment: consequential decisions remain with the people who own them.

The benefit is narrower and more useful. Intent and accountability survive while current evidence changes the route to the outcome.

What to inspect

Ask the authority that can prove the current claim:

1. Does the Task still express the accepted outcome?
2. Which Assignment has the one active owner?
3. Is uncertainty represented as a Scout, decision or dependency?
4. Did new work become a distinct Task for a concrete reason?
5. Which native artifact proves delivery?
6. Which action still belongs to a human?

Organizational attention explains how the Fleet wakes the right owner and returns material decisions without turning a person into its polling loop.