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.