Write the escalation path for an agent that gets stuck

Agent: {{what it does}}. It runs {{how often}}, and when it stalls today the result is {{what actually happens — usually nothing}}. Humans available: {{roles and hours}}.

Define the escalation path.

First, list the distinct ways it can get stuck: tool failure, ambiguous input, low confidence, a policy boundary, a loop, and running out of budget or time. These need different handling — don't collapse them.

For each: what the agent tries first on its own and how many times, when it gives up, who it goes to, through which channel, and what it must include in the handoff so the human doesn't have to reconstruct the run.

Then the parts everyone forgets: what the end user sees while it's escalated, what happens outside working hours, what happens if nobody responds within {{window}}, and how the resolution gets back into the system so the same case escalates less next time.