Looping
The agent keeps making new steps inside the same failed frame. The missing mechanism is not effort. It is boundary redefinition.
AI Agent Failure Recovery
Your agent is not failing because it lacks another rule.
Many agent failures are recovery failures. The agent can continue producing actions, but it cannot recognize when the original frame has broken.
Hallucination
Hallucination is often a boundary problem. The agent treats uncertainty as permission to continue instead of a signal to stop.
Tool failure
More tools can make weak recovery worse. The agent gains more ways to act without knowing when action should become adjustment.
The recovery diagnosis.
When an agent fails, do not only ask what instruction is missing. Ask what signal would make the agent stop, review, and change itself. A good failure recovery loop changes the next action by changing the internal frame first.
This page exists for builders who work with long-running agents, browser automation, tool use, coding agents, and workflow copilots that must remain coherent after the first plan fails.
A compact checklist.
Did the agent identify the failed assumption? Did it name the boundary? Did it simulate a safer alternative? Did it ask for missing input? Did it change strategy, or only add more output?