The situation
A contractor leaves on Friday. Their seat is removed. On Monday you find the coding agent they set up still has a token with write access to the payments repository, because the agent ran on a shared service key that nobody thought to rotate.
Three ways agents get access
| How the agent gets access | Who it can reach | What happens when someone leaves |
|---|---|---|
| A shared service key | Everything the key can reach | Nothing, until someone rotates the key |
| A key per team | Every repository the team has | The team's key keeps working |
| The engineer's own access, per repository | Only what that person may touch | Their agents lose access with them |
How to set it up
- No shared keys. Agents run on each engineer's own machine with that engineer's own credentials. Write it down as a rule every agent works under.
- Grant by product and repository. A developer building the web app gets the web app's repositories, not billing's.
- Grant agents per person. Decide which agents each person may use, product by product.
- Make risky steps stop and ask. Deploying, changing infrastructure or running a penetration test waits for a named person's approval.
- Record every grant and removal. Who got access, who removed it, when and why, in a record nobody can edit.
Check yourself
- If someone left today, how many keys would you need to rotate?
- Can you list, for one repository, every person and agent that can write to it?
- Which agent steps can reach production without a person saying yes?
With Wakalix
Agents use each engineer's own access, never a shared key, and Wakalix never runs the model. Access is granted per person, product and agent, risky steps wait for a lead's approval, and every grant and removal is on the record, with a quarterly access review to export. See Governance and Security.