Operating notes

Four things the team that installs the skill will meet in practice: agents that run for minutes, the description that decides when the skill fires, narrowing a skill to a tag, and what happens when no agent matches.

Long-running agents

A script call returns when the agent finishes. Set the timeout generously, and for agents known to run for minutes, have the skill return the conversation id and the GUI link so the person can follow the run there. The streaming endpoint is available to a script that wants to print progress.

Testing when the skill is used

The description in SKILL.md is the only thing the assistant reads to decide whether a request is for the company's agents, before Search2o is called. Too broad, and general requests get sent to search; too narrow, and requests that should reach an agent never do. Try a handful of requests that should use the skill and a handful that should not, and adjust the description's examples until both behave.

Tag scoping

search accepts a tag. A company that wants narrower triggers can install two skills — support agents, finance agents — each searching its own tag.

Refusal is a feature

When Search2o returns nothing, the assistant answers from general ability. Search2o never guesses which agent to run, and the person still gets an answer. See How matching behaves.