Reading your run log: what your agent did and why
Every run leaves a record of the jobs it saw, the choices it made and the tokens it used. How to read it and what to fix.
The run log is how you stay in charge of an agent you are not watching. Each run writes down what it searched, what it decided and what it cost. If you only look at one page when something seems off, make it this one.
What a run records
- When it ran and whether a schedule or a person started it.
- Searches, one entry per keyword.
- Decisions, with a short reason for each job it read, drafted for or skipped.
- Tools used, such as search, reading a job post or preparing a proposal.
- Outcome and tokens, including runs that found nothing.
Skipped and stopped runs are information too
A run that skipped because no keywords were set, or stopped because it hit its time limit, tells you what to change. Silence is never a good sign; the log makes it visible.
Questions the log answers
- Why did nothing happen last night?
- Why did it apply to that job?
- Which keyword is finding the good jobs?
- How many tokens did this agent use this week?
Turn what you learn into settings
If a keyword keeps returning jobs that get skipped, replace it. If good jobs are filtered out, relax a filter. If runs are slow or costly, run less often or judge fewer jobs per run.