12 · Several agents on one task

Splitting work across several agents sounds like a team, and teams sound good. But each extra agent is another loop with its own prompt, its own context and its own ways to go wrong, plus the work of passing results between them. Most tasks are better done by one agent with good tools.

A few situations do push past one agent. The material won't fit in one context window (reading 50 files to answer a question). The work splits into independent parts that can run at the same time. Or parts of the job need very different tools and instructions that would confuse one prompt.

When one agent struggles, look at its context before adding a second. If it's missing the schema, a decision someone made or the command that runs the tests, a new agent will be missing them too. Better context for each agent often helps more than adding agents.