What broke when I ran a leaderless agent swarm with no orchestrator — 7 failure modes and fixes

LostHumanity67 · reddit · 2026-11-02

The author ran DreamTeam, an open-source (MIT) leaderless agent swarm with a shared task board, git workspace, and agents that write their own MCP, on four long tasks with checkable answers. The most valuable output was the failure list:

Planner hoarding research: agents idled 10–40 minutes. Fix: planning prompt creates independent tasks first and delegates research; a hint fires after 8 steps or 10 idle minutes.
Final report depending on optional exploratory tasks, so it could never start. Fix: final task depends only on essentials; assignee may drop extra dependencies.
Runaway wrap-up: told to wrap up, an agent launched another "quick" 16-process run. Fix: wrap-up mode outright refuses new background jobs.
Truncated thinking: an architect produced 30 minutes of thinking cut off by maxtokens. Fix: double the output limit up to a ceiling after truncation.
Junk in git history: temp test files and nested repos got committed. Fix: commit a new file only if it survives the agent's next command.
Networked agent fetched an upstream version with the bug already fixed. Fix: a no-network session mode.
Peer review checks completion, not correctness: the swarm "beat" a published table on 57 data points that was simply under-converged; only an outside check (hidden benchmark, independent recomputation) catches this.

These map to the MAST taxonomy of multi-agent failures (Cemri et al., arXiv:2503.13657): specification, inter-agent misalignment, verification.

Original post →

What broke when I ran a leaderless agent swarm with no orchestrator — 7 failure modes and fixes

More from coding & agent