Harness, Part 7: Channels & Triggers
Your agent runs while you are away: one job, one schedule, one log - automation without frills.
What this is about
Everything your agent can do so far needs you at the terminal. That is the definition of a conversation - but automation is its opposite: a job runs at a time and leaves a trace. Those three words are the entire blueprint of this part, which belongs to the extensions from Block II: a one-shot mode that works through a job without a single follow-up question, a schedule that kicks it off, and a log that shows you afterwards what happened.
The delicate point is not the technology but a consequence: in one-shot mode, nobody sits in front of the confirmation gate from Part 5. The answer is not to switch the gate off - but to tighten it: what would be a question in conversation becomes a refusal in an unattended run. Your agent would rather do nothing than do the wrong thing.
At the end of this part, this works: your agent has completed a task while nobody watched - and you can read up on how.
How the build works
As before: copy the prompt, hand it to your coding agent, answer its questions. The prompt reads harness-doku.html and requires the state from Part 6.
This section belongs to the Pro tier. You get access to full code, deeper explanations, and rabbit-hole material.
See tiers →It works when
A run starts and ends without anyone at the terminal - once kicked off by hand, once by the schedule. The log tells you afterwards what happened. The more often your agent runs, the more expensive long conversations and jobs become - time for the keyword "context-frugal": Part 8: Context Compaction.