Your agent is creating evidence debt

An agent completes a task, reports that the checks passed, and moves on. The change works. Nobody asks for more detail.

Three weeks later, another person needs to understand why the change was safe. The run log is gone, the screenshot has no build identifier, and the summary says only that all tests passed. The patch is still there, but the reason it was accepted has disappeared.

That is evidence debt. The workflow took the benefit of a fast decision without paying the cost of making that decision inspectable later.

Logs are not automatically evidence

Agent systems produce plenty of material: tool calls, token traces, terminal output, screenshots, model explanations, and chat transcripts. Volume can hide the problem. A pile of run data is not the same as evidence for a decision.

Evidence has to answer a question. Which version was tested? What behavior was observed? Which policy applied? Who or what accepted the result? Can the result be tied to the exact change that reached production?

Consider a deployment where the agent records a green test run and a screenshot of the homepage. If the tests ran against one commit while the screenshot came from an older deployment, both artifacts are real and the conclusion is false. Without identity and provenance, evidence can decorate a decision without supporting it.

Agent-written summaries need the same skepticism. A summary is useful for navigation, but it is still a claim made by the system that performed the work. “No regressions found” should point to the checks, comparison, and scope that justify it. Otherwise the next operator has to repeat the investigation.

Debt compounds when the system changes

Weak evidence may seem harmless while the original operator remembers the run. That memory acts as an undocumented index. It disappears when the person changes teams, the model changes, or several similar runs blur together.

The cost surfaces during a rollback, an audit, or a regression investigation. People rerun old tests against new dependencies. They infer intent from code. They compare screenshots by eye and hope the environments match. Work that felt fast at acceptance becomes expensive when the decision has to be reconstructed under pressure.

The debt also limits learning. If a team cannot connect accepted changes to their evidence, it cannot tell which checks predict good outcomes and which ones merely produce green marks. Every evaluation starts from anecdotes.

Preserve a receipt, not the whole run

The answer is not indefinite storage of every token and shell line. Most raw traces are noisy, expensive, and full of context that should not be retained. Some contain secrets or personal data. Evidence collection needs limits too.

A durable receipt should be small enough to inspect and specific enough to replay. For a software change, it might identify the input task, proposed diff, tested revision, check results, decision, and rollback reference. A customer support workflow might retain the policy version and approval record without keeping unrelated customer data in the agent’s reasoning trace.

Match the receipt to the lifespan and blast radius of the action. A private draft that will be discarded tomorrow does not need the same record as a schema migration. A reversible internal change may need less than a payment or a public claim. The point is not universal retention. It is deliberate retention.

The receipt also needs an owner. If nobody is responsible for whether it remains readable and linked to the change, the storage bucket becomes another form of debt.

Pay at the point of acceptance

Choose one recurring agent action that can affect production or another person. Define the minimum receipt required before that action is accepted. Make the workflow capture identifiers and evidence automatically where it can, then ask the approving authority to record the decision rather than relying on the agent’s recap.

Test the receipt by giving it to someone who did not watch the run. They should be able to identify what changed, why it passed, and where to start if it must be reversed. Any question that sends them back to a vanished chat is unpaid debt.