Agentic Goals

Turn an explicitly requested outcome into durable, evidence-backed pursuit state. Do not create a goal implicitly from an ordinary task.

Pursue the goal

Confirm the objective in outcome language. Record measurable success criteria, constraints, authority boundaries, budgets, and an initial next action.
Create native goal state when the runtime exposes goal tools and the user explicitly requested goal pursuit. Otherwise maintain the JSON contract locally in the authorized workspace.
Select the narrowest execution protocol: use agentic-loops for adaptive serial work, agentic-graphs for dependencies and parallelism, or council for consequential judgment.
After each meaningful action, append evidence with source, result, affected criterion, freshness, and explicit supersedes lineage when replacing an earlier result. Use only the one current chain head per criterion to decide progress, so stale passes cannot mask newer failures. Update the next action from current evidence.
Resume by reading existing goal state first. Preserve verified progress and do not replay work unless evidence is stale or invalidated.
Mark complete only when every success criterion is explicitly verified and no required work remains.
Mark blocked only after occurrence records prove the same external blocker has prevented meaningful progress for at least three consecutive goal turns ending at the current turns_used, and recorded evidence shows no safe alternative remains. Before that threshold, keep the goal active and pursue safe alternatives.