Design concept by Krishna Bhupathi. Aftercast is not a real product, and this incident is invented.

See the studio

Report

AC-0417

Occurred

2 Sep 2026, 02:14 UTC

Agent

Refund agent, v3

Operator

Online retailer, name withheld

Run length

41 min 02 s

Status

Final report

The refund agent paid 312 orders twice, and reported success.

The refund agent paid 312 orders twice, and reported success.

This report replays the run from its Aftercast recording. Drag the playhead to see what the agent saw, what it did, and the moment it went wrong.

Drag the track, or focus it and use the arrow keys. Page Down jumps to the next finding.

An engineer working alone under a desk lamp in a dark office at night

02:14 UTC

One engineer on call. The dashboard showed every refund as a success.

Summary

Agents rarely crash. They keep going, confidently, in the wrong direction.

Agents rarely crash. They keep going, confidently, in the wrong direction.

For 41 minutes the agent did what it was asked: refund customers whose orders arrived damaged. When the payments service slowed down, one refund looked like it had failed. The agent tried again. It had already worked.

Every log line said 200 OK. The agent's own checkpoint said “all successful”. It took a finance alert, not an engineering one, to stop it.

1

Sequence of events

What happened, minute by minute

What happened, minute by minute

Time

Event

00:00

The run starts with 1,140 refund requests in the queue.

06:14

The payments service slows to 9.8 s per call. The agent's tool gives up after 8 s.

06:24

A refund times out. It has in fact gone through.

06:31

The agent retries and refunds the same order a second time.

14:50

The agent's checkpoint: 412 refunds processed, all successful.

38:40

Finance flags refunds at 3.1× a normal day.

41:02

The on-call engineer stops the run. 312 orders were refunded twice.

2

Findings

Three things went wrong. None of them was the model.

Three things went wrong. None of them was the model.

F1

The timeout hid a success

The payment went through. The tool stopped waiting at 8 s and told the agent the call had failed.

F2

Retries had no idempotency key

Each retry was a brand-new refund, so the payments service had no way to know it was a repeat.

F3

Success was counted, not checked

The agent counted calls that returned 200. Nothing compared refunds paid with refunds owed.

3

What the agent believed

Its confidence never dropped below 0.9 while the error grew.

Its confidence never dropped below 0.9 while the error grew.

Confidence is the agent's own score for each step. Orders paid twice are counted from the payments ledger after the run.
An inspector in a hard hat reaching up to the underside of an aircraft wing

Why a recorder

Air accident investigators replay the flight recorder before they change a rule. Agents that spend money deserve the same.

4

Recommendations

Record every run. Replay before you trust it. Turn findings into guards.

Record every run. Replay before you trust it. Turn findings into guards.

This is what Aftercast does. It sits between your agent and its tools, so nothing in your prompts or code changes.

R1

Record

Every prompt, tool call, response, cost and confidence score, with the exact time it happened.

06:24.118

payments.refund

order

A-56230 amount 52.00

timeout

8000 ms

R1

Record

Every prompt, tool call, response, cost and confidence score, with the exact time it happened.

06:24.118

payments.refund

order

A-56230 amount 52.00

timeout

8000 ms

R2

Replay

Scrub any run like the one above. See what the agent knew at each moment, not just what it logged.

run

AC-0417 at 06:31

paid twice

1

confidence

0.94

R3

Guard

A finding becomes a rule that stops the next run before it repeats the mistake.

guard

one-refund-per-order

when

payments.refund

block

repeat within 24h

A cardboard parcel sealed with red Fragile tape

5

From the interviews

“The logs said 1,140 refunds, all successful. Both numbers were right.”

“The logs said 1,140 refunds, all successful. Both numbers were right.”

On-call engineer, name withheld

Other reports

More runs, replayed

More runs, replayed

Request access

Put a recorder on your next run.

Put a recorder on your next run.

Aftercast is in private beta with teams running agents in production.

Aftercast is a design concept by Krishna Bhupathi. The product, the incident and the people are fictional. Photos from Unsplash.

bhupathi-studio.framer.website

Create a free website with Framer, the website builder loved by startups, designers and agencies.