After a tool fails, an approval expires, or a user corrects the plan, the agent should not simply retry. It needs a recovery prompt that summarizes what happened and chooses the next safe route.
Exception memory records what happened. The recovery prompt converts that record into a concise instruction for what the agent should do now, what it must avoid, and what evidence it should preserve.
It should name the failed action, the blocking condition, the user correction or fallback, and the proposed next route. That keeps the agent from losing context and repeating the same brittle step.
Summarize what failed and switch to a safer path before retrying.
Preserve the draft and explain what happens when the approval window expires.
Convert the correction into a routing rule before continuing.
State the attempted action, the blocker, and the result. The agent should not infer this from scattered chat history.
Include the user correction, failed tool, expired approval, stale context, or human handoff outcome that changes the next route.
Name the path that should not be repeated: a tool, a claim, a public action, a source, a timing assumption, or an unsafe shortcut.
Choose a safe continuation: draft, retry with constraints, ask a narrower question, route to a human, or stop and report.
They should appear inside the workflows where agents actually lose context: text, browser work, publishing, and handoff.
Use text message AI assistant flows to recover from ambiguous or delayed replies.
Pair with computer-use cache when browser actions fail or require safer context gathering.
In agent website workflows, recovery prompts should mention broken links, stale deploys, and release gates.
Summarize what a person did so the agent resumes with the correct context.
Is this just retry logic?
No. Retry logic repeats an operation. A recovery prompt changes the agent's context and constraints before the next operation.
Should users see recovery prompts?
Often yes, especially when the next step changes risk, timing, or approval behavior.
Where does Supers fit?
Supers is relevant for personal AI agents that span text, browser automation, publishing, and human escalation.
Recovery prompts make exception memory actionable. They tell the next agent attempt what happened, what changed, what to avoid, and how to continue safely.
Use recovery prompts to preserve the failure context, protect the user, and keep the next action aligned with what actually happened.