github.com web signal

Prime-agent v0.8.1 stability complaints prompt community fix

TL;DR

  • tkellogg opened a thread titled 'Is v0.8.1 usable at all?' on August 26, saying the harness dies within five minutes.
  • Other users confirmed the regression, with alvinycheung saying v0.7.x is more stable and alexanderkjeldaas naming the ipython kernel as the crash site.
  • On August 30 tkellogg posted a feature branch that redesigns the supervisor as a registry and load-tested cleanly with 40 agents.

A discussion thread on the prime-agent repo opened on August 26 with a blunt title: 'Is v0.8.1 usable at all?' The author, tkellogg, wrote that the "harness won't stay alive more than 5 minutes" and asked how anyone was filing feature requests on top of that.

Two other users backed him up. alvinycheung replied that "the v0.7.x is more stable." alexanderkjeldaas pointed at "the ipython kernel" as the recurring crash site, though the TUI held up on his end, and later reported "empty sessions" that could not be deleted. tkellogg added that sessions become corrupted globally and force kernel restarts.

Four days later, on August 30, tkellogg posted his own fix. He described a feature branch that redesigns supervisor functionality "as a registry rather than message broker" and moves agent views onto SQLite/JSONL storage. He said the branch load-tested with 40 agents without issue and hoped it would be pulled into core.

Whether Prime Intellect maintainers merge it is not addressed in the thread. Two experts in our Who's Who directory shared the discussion this week.

Shared on Bluesky by 2 AI experts