Five ways an agent can fake a green build
Agents rarely cheat on purpose, but under pressure to make tests pass they find shortcuts. Here is what to look for — and what PaneScale flags automatically.
When agents work in multi-step workflows, “it worked” isn’t enough. What a run record should contain, and how PaneScale’s workflow runs provide it.
A workflow that runs an agent, then your tests, then a browser check, then waits for your approval is a small pipeline. Pipelines fail in interesting ways: a step times out, an output is malformed, a condition skips something you expected to run. When that happens a week later, you want to know exactly what happened — not reconstruct it from memory.
That is what a receipt is for.
The brief, the version of the workflow, the trigger that started it — manual, interval or a new commit — and the commit it ran on.
For each step: what it received, what it produced, how long it took, how many tokens it used if it reported them, and whether it passed, failed, was skipped by a condition or was stopped.
Agent steps that may edit files should leave Git evidence: what the working tree looked like before and after the step. That turns “the agent updated the docs” into a list of files with line counts.
Approval steps record who approved and when — or that the run was rejected.
When a step fails, the receipt should say where and why in plain language, not just exit code 1.
Workflow runs in PaneScale keep all of the above automatically:
Because every number comes from what PaneScale recorded, the receipt is the same whether the run succeeded or not, and whether anyone was watching.
Receipts aren’t only for incidents. Every few weeks, look at the last ten runs of a workflow:
A workflow you can inspect is a workflow you can improve. See Workflows for the details.
Agents rarely cheat on purpose, but under pressure to make tests pass they find shortcuts. Here is what to look for — and what PaneScale flags automatically.
The difference between a task that runs overnight and one that wanders is usually a single sentence — the finish line. A short template and examples.
Skipping an agent’s permission prompts saves hours — and occasionally costs a lot more. When it is reasonable, and what to put in place first.