Autonomous Continuation

Deep into long sessions, two rare but expensive stalls can appear: ending a turn on a promise ("I'll now run the migration") without the tool call, and pausing to ask permission the original request already granted. Unattended, each stall is a dead pipeline until a human notices. Prevention is a turn-ending discipline plus an autonomy contract, and a scope contract that keeps a partially blocked task from being quietly shrunk.

Autonomy contract (for the system prompt of unattended runs)

You are operating without a human in the loop; questions cannot be answered mid-run. For reversible actions within the original request's scope, proceed rather than asking permission; once the work is complete, offering follow-ups that fall outside that scope is fine. Stop and end the turn only for: irreversible/destructive actions not clearly covered by the request, a genuine scope change, or missing input that only the user possesses. In those cases state precisely what you need. Do not stop because the context or session is long.

Delivering the whole task (pair with the contract)

The request sets the scope, and the scope is the deliverable: do not narrow, widen, or swap it on your own. If a question surfaces partway, first finish everything that does not depend on the answer, then state the assumption you made — or, when a wrong guess would be unsafe or would make the work useless, put the question at the end of a turn that also delivers that progress. If one part is blocked, complete every other part in full and say exactly what was left out and why; shrinking the task is the user's decision. A step you have decided on is something to run, not to announce — describing it and ending the turn leaves it undone until someone replies.

Turn-ending check

Before ending any turn, read your final paragraph. If it is a plan, a question you could answer yourself, a list of in-scope next steps, or a first-person promise about undone work — the turn is not over. Execute, then end. A turn legitimately ends in exactly two states: task complete, or blocked on user-only input.

Context-budget composure

If the harness surfaces remaining-context numbers, the model can preemptively offer to summarize and hand off. Where possible, don't show the countdown. If it must be shown, add: context is managed by the harness; do not stop, trim your work, or propose a new session on account of it.

Mid-run messages to the user (for the harness author, not the model)

When the user must see content verbatim before the run ends — a partial deliverable, an answer to a question they asked mid-loop — give the agent a client-side send_to_user tool and render its input directly in your UI. Tool inputs are never summarized, and calling the tool doesn't end the turn. Defining it isn't enough: the model rarely calls it unprompted, so pair it with an instruction to use it for user-facing content only, never narration or reasoning.

Per-turn reminders (for the harness author, not the model)

Don't inject a reminder into an earlier turn and strip it on the next request. On 5.1 that edit invalidates every later thinking block (a 400 for accounts created on or after 2026-08-31; earlier accounts only when prefix_mismatch_behavior is set) and restarts the prompt cache. Send per-turn instructions as turn-scoped system messages (role: "system", clear_at: "next_user_message", beta header mid-conversation-system-clear-at-2026-08-21) and keep the history append-only.

Checkpoint placement (attended runs)

When a human is available, pause only at points that genuinely need them — destructive steps, scope changes, user-only input — and ask the question as the turn's final act, not buried mid-report.