← Episode prep
3 AImigo · Episode prep
Infrastructure for agent-driven development
Give several tasks to agents in the evening. By morning you have proposed changes, test results and new questions. Who checks the work, allows a merge or stops a dangerous action? What happens when thousands of agents are doing this at once?
Questions for the discussion
- How do you control the flow of work and keep its history when an agent fails?
- What happens to CI, version control and merge queues as the number of attempts grows?
- How do you verify the result and check the verifier itself?
- How do you grant and revoke an agent’s permissions?
- How do you calculate the cost of an accepted change, including failed attempts, checks and rollbacks?