agent run history and reports

Agent run history and reports are the paper trail that lets you answer plain questions about AI assisted work weeks later. What ran on this task? What did it change? Did the checks pass, and were they still valid by the time someone merged? How much did it cost? Without a record you answer from memory and scrollback, and both are unreliable. With a record you still have to read it carefully, because histories can look more complete than they are. Entries can be untagged, lists can be truncated for display, usage can be missing because a tool never reported it, and a name on a line can be a label rather than a verified identity. This page covers what a useful run history contains, how to read gaps without inventing numbers, and how Liquid Brain presents reports and usage coverage.

What a run history should capture

A run record answers a handful of questions every time. Which task did it belong to and what was the objective? What was attempted, when, by which agent? What checks ran and what did they return? Which paths changed, and what became of the commit? What did it use, to the extent the tooling reports usage? And what did a human decide afterward? Keep it per attempt, not per task, because a task with three tries and one success is a different story from one that worked first time. Keep the record next to the work, so it travels with the item and is found by the same search. The point is not surveillance of the agent. It is the ability to debug your process: to see that a certain kind of task always needs two attempts, or that a check keeps passing on work you later reject.

Reading coverage: unknown is not zero

Reports are summaries, and every summary has a denominator. Ask what it covers. An interactive terminal session is not the same as a managed run, so a history of managed runs will not include what you did by hand. Usage figures are the clearest trap: if a tool did not report usage, the total is not low, it is incomplete, and presenting it as a smaller number misleads. The same goes for invoices and account wide spend, which a task level record cannot reconstruct. A saved report is an observation from an earlier moment, while a live query reflects now, so the two can legitimately differ. And identity needs care: a history line saying a certain agent did something is a declared tag unless authentication backs it. Good reports disclose omissions explicitly, and good readers look for the disclosure first.

Reports and run history in Liquid Brain

Liquid Brain keeps recorded attempts, outcomes and reported usage with the work [C4]. Reports show decisions waiting for attention, recorded work history and dated items, with missing coverage disclosed [C7]. History is grouped by recorded week, actor and project, with untagged entries and source details visible [C7]. Management also holds a run history of managed harness attempts, which excludes ordinary interactive terminal sessions [C4]. You can inspect recorded usage and its coverage instead of reading missing numbers as zero [N06]. Advisory project thresholds do not enforce a hard spend cap [N06]. Evidence freshness shows whether recorded evidence still matches its files [C6], and the optional project digest lets you revisit recorded decisions and literal mentions inside the app [N07]. Refresh rereads recorded data and does not rerun any agent.

QUESTIONS

What should an AI agent run log include?

The task, the attempt time and the agent, the checks that ran and their results, the changed paths, the commit outcome, reported usage where available, and the human decision afterward. Keep it per attempt rather than per task, so retries stay visible and a pattern of repeated failures can be spotted later.

Why are some usage numbers missing in my reports?

Not every tool reports usage, and interactive sessions or invoices can remain unknown. Liquid Brain shows recorded usage together with its coverage and treats missing figures as unknown, not zero [N06]. It makes no whole account invoice claim [C4], so use it to compare tasks, not to reconcile a bill.

Are actor names in agent history reliable?

Treat them as labels. In Liquid Brain Reports, a declared actor tag is distinct from an authenticated identity, so use it as provenance to inspect rather than proof of who acted [C7]. If you need verified attribution, that has to come from somewhere other than a name on a history line.

RELATED GUIDES

• How to read Liquid Brain reports and provenance
• How to read project progress in Liquid Brain Build
• Preview: a new Gantt planning experience for Liquid Brain

RELATED HARNESS PAGES

• human review of agent output Human review of agent output: what to inspect before accepting AI changes, and how recorded outcomes make the review honest.
• evidence gates for AI work Evidence gates for AI work: separate proposed from approved execution, run checks, and record outcomes so acceptance rests on evidence.
• scheduled agent tasks Scheduled agent tasks done carefully: explicit prompts, readiness and authority checks, and why a started run is not a finished one.