Imbue, the AI research lab, published part two of a case study on their engineering management tool mngr, detailing how they ran 100 Claude-powered agents simultaneously for testing.
The setup pushes into territory most teams haven't explored yet. Running a handful of AI agents is straightforward. Running a hundred at once for coordinated testing introduces real engineering problems: API rate limits, cost control, making sure agents don't collide on shared resources, and validating that outputs stay consistent when you can't manually review each one.
Imbue's mngr product uses AI agents to handle engineering management tasks, and Part 2 of their series focuses specifically on the testing infrastructure required to validate agent behavior at that scale. The practical question they're tackling is one that matters as more companies move from "one developer, one AI assistant" toward deploying fleets of agents across their workflows: how do you actually QA a system where the workers are non-deterministic language models?
For teams already running multi-agent setups with Claude, the case study offers a look at what the next order of magnitude feels like. The full writeup is on Imbue's blog.