Sergey Kuhne describes getting overwhelmed as coding-agent sessions multiply and each reply demands another decision. His experiment puts tasks and decisions in a Git-backed queue, so he can focus on one item without holding the rest in his head.
- An agent researches tickets and builds those marked safe; sensitive changes go to a branch for approval. Kuhne retains production deployment.
- An owner inbox collects questions and reviews requiring his judgment. A generated page presents the context, while Git keeps the record.
- He decides on one item and returns it to the queue before loading the next. He says the process is still exhausting, but less distracting; this is one person’s experience, not a measured productivity result.
The 133-comment thread on Hacker News adds a separate concern: keeping track of tasks does not eliminate the work of reviewing what agents produce.
What the thread adds
- sollewitt — says large code reviews are the real bottleneck and suspects serial work may be faster than parallel work when review time is included.
- eterm — limits concurrent tasks to two or three and takes breaks rather than automatically starting another session while waiting.
- stillpointlab — reports that Markdown tracking helps, but reloading context and facing accumulated reviews remain exhausting.
- mr_roboto — describes a contrasting workflow: batching many agents and reviewing afterward, while noting that some software is better suited to it than others.
Handles are pseudonymous, HN publishes no per-comment scores, and ordering follows HN’s ranking. These comments are a slice of the discussion, not a consensus.