What it is for
An AI agent is the answer to "the summary is fine, but it does not speak our language". Every AI task in meulog — the period summary, reading one error, triage, the customer-facing text — ships with a prompt written in the code. An agent replaces that prompt for one task, without touching the project.
Creating one used to mean writing a Ruby class. Now it is a form, and the default prompt of every task is right next to the empty field: the gesture is editing what already works, not writing from a blank page.
What each field decides
- Task is where the agent enters. One list, closed: it is the call sites that exist in meulog, and a task nobody calls would be an agent that never runs.
- Prompt replaces the one in the code. Empty means the default one, which is shown below the field.
- Glossary is the vocabulary of the house — rejection codes, product names, what an abbreviation means. It is prepended to the prompt.
- When it applies routes: project, error class, app, and the minimum number of occurrences. The expensive model takes what has volume, and the order sends the rest to the cheap one.
- Ceiling and price are the daily budget of this agent, in cents, and what a million tokens costs. Zero means no ceiling.
Rehearse runs the agent against a real error and shows the exact question that left here and the answer that came back. It spends real tokens, and that is the point: it does not deliver anything — no channel is told, no issue is opened, no triage is stored. It is the difference between "I think this prompt got better" and "this prompt got better".
Improving the prompt
Improve prompt asks the model to rewrite its own prompt. The proposal lands inside the field and is not stored: a button that rewrites and saves by itself changes how every error of the night is read without anyone having read the change.
Did it get better?
The thumbs on the rehearsal answer are the only way to compare two prompts without guessing, and the score belongs to the version: changing the text carries the tally to the version that leaves, and the agent starts at zero. The history below the field shows what each text was worth while it was in force.
What it costs
The top of the screen shows the whole account's spend over 30 days, one line per spender — including the plugin that answered with no agent at all. The daily ceiling for the account lives in Settings: seven agents each inside their own ceiling add up to the invoice nobody saw coming.