Sheet B-05 / Builds

Detekta

AI agents will tell you a job is done. Detekta checks whether it really is.

It runs the agent in a clean sandbox, looks at what actually changed, and gives a verdict backed by evidence. It's for teams building AI agents.

I set it aside for a while because I didn't know enough about AI agents yet. Then it hit me. Building and running AI agents is the sexy part. Checking their work is the boring part. Detekta does the boring part.

It hasn't launched yet.

Started
Aug 2026
Status
Building
Kind
Testing

The job

Create a customer named Jane.

The agent says

Done. Jane is created.

Detekta checks what changed

  1. One matching row is in the database: passed
  2. Looking Jane up returns Jane: passed
  3. No duplicate customer exists: failed

Not done. Jane was created twice.

Fig. 1 A drawing of the idea. The job and the result are made up.

What it's made ofYou only ever see the top layer

  1. 1A verdict backed by evidencewhat you see
  2. 2Checks what really changed
  3. 3Runs the agent in a clean sandbox