Every tool def this package ships, in the order an agent should read them: operate a
running instance, then author a definition (see WorkflowToolDef).
Each tool's own description is written for an agent, not a human: what it does, when
to use it, what not to use it for. The write tool spells out that it's the only way to
advance state — without that line, an LLM looks for a complete_activity or
set_stage tool that doesn't exist. Read tools cap their output rather than burn
context on the rest.
Every tool def this package ships, in the order an agent should read them: operate a running instance, then author a definition (see WorkflowToolDef).
Each tool's own description is written for an agent, not a human: what it does, when to use it, what not to use it for. The write tool spells out that it's the only way to advance state — without that line, an LLM looks for a
complete_activityorset_stagetool that doesn't exist. Read tools cap their output rather than burn context on the rest.