HN Explorer

Ask HN: What broke when you added the second agent?

1 points by druhinbala ↗ HN
Most multi-agent write ups are about quality - which model to use and why, more agents and generating better answers. I am interested in the failures that happen.

If you run more than one agent against shared state or shared tools, what actually went wrong?

1. Did two agents do the same work, or different work that conflicted? 2. Were your agents copies of each other (homogenous agents), or specialized with different roles (heterogenous)? 3. What did you add to stop or prevent the failure?

Anthropic's multi-agent research found that agents are low-variance: 18 of 30 independently created a git branch with the identical name. That suggests the common failure is correlated duplication rather than disagreement. Does that match what you have seen, or do yours diverge?

3 comments

[ 1.5 ms ] story [ 11.0 ms ] thread
verdverm
they are so reliably inconsistent, but markdown engineering goes a long way towards improving this (eg. we have zero branch name issues like Anthropic claims)

the biggest issue I've seen are these inconsistencies, mistakes, bad assumptions, or side quests compounding with multiple agents and no human-in-the-loop

spent some hours yesterday refining my orchestrator prompt, so much better - tl;dr use the default implementation agent unless the user specifically asks you to use the big/hard multi-agent workflow (which has agent-human loops and break points between, for each phase)