Infrastructure for Agent Development
Episode hosts
What we discussed on the recording
Alexander Polomodov, Evgeny Sergeev and Aleksey Litvinov discuss how a stream of agent changes loads development infrastructure. Agents need working copies, environments, storage and checks as well as tokens. More changes can shift bottlenecks to CI, review and integration without delivering value to users any faster.
Parallel agents can alter related parts of a system or duplicate work. Queues help regulate the flow, while explicit rules can expose semantic conflicts. Evgeny describes an inbox for agent questions and results, allowing a person to respond at their own pace. The hosts note that a queue can also become a constraint.
The hosts compare checks before and after merging and discuss risk-based review automation. Scion and Symphony illustrate different approaches to storing changes and coordinating tasks. These solutions reflect the needs of particular teams; the principles and trade-offs are more transferable than the entire infrastructure.
The conversation closes with agent permissions, isolation and secret protection, then the cost of an accepted change. That cost includes retries, checks, human work and fixes as well as tokens. Faster code delivery is still constrained by the time needed to collect feedback from users.