Agent Engineering HandbookGitHub

Idea 02 of 19 The nineteen ideas

Design tiny two-actor end-to-end tests that catch whole classes of failure

On this page

Theo's chat test caught failures early

Theo recalls a Twitch test that opened two Playwright browsers with bot accounts already signed in. One bot sent a chat message, while the other checked that the message appeared and rendered. He says this simple test caught failures earlier than almost anything else in that system.

Test that the other person receives it

Choose the smallest path where one actor produces something and another actor must receive it. Encode that path as an end-to-end test using authenticated browser contexts for the participating actors. Assert that the receiving side actually sees the result when shallow checks keep missing the same integration failure.

Separate checks can miss broken interactions

This pays off for chat, notifications, or other multi-party features where component checks pass while integration breaks. The important situation is a failure across the actors' shared path that isolated checks keep missing. Do not select a two-actor test merely for its shape when your failing behavior does not involve that interaction.

The recalled test has no implementation here

Theo is recalling a test from a past employer, rather than supplying a checked-in implementation here. The video does not prove the separately documented details of video retention, tracing, or retries. Use the cross-system assertion as a pattern when it matches your failure, without treating those details as demonstrated.

Watch at 03:46 · Home · Agent Engineering Handbook, September 9, 2026