Repository navigation
Claude Long Running Autonomous Build of podbench #6
gilesknap
announced in
Announcements
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
podbench build session stats
The initial build of this was run in a single workflow using an artifact plan developed on a phone with Claude.ai.
First release was commits to main and fully worked against my cluster that I gave it access too for the workflow.
Subsequent PRs refined and got fully working against DLS cluster.
Subagent work: 10 workflows, 29 agents
About 3.4 hours of agent compute, heavily overlapped. The four big phases each ran 4 to 6 agents concurrently.
What came out
Published artefacts: a multi-arch container image and a Helm chart.
Where the money went
Phase 0 spikes were about 15 percent of spend and prevented the most rework. They falsified five load-bearing assumptions before any product code existed. One of those was that kubectl debug copy-to strips all labels, which would have made the headline demo unbuildable as specified.
Review agents were the best marginal spend. Every batch shipped defects that its own authors had reported as done: a crash in the exact code path it was written for, ten wrong documentation claims, and a quickstart that could not work.
There were 453 green unit tests at a point when the tool could not open a single ssh connection. Four bugs were reachable only by running against a real cluster:
All reactions