More agents are better than one

The argument

Multi-agent systems have an intuitive appeal: one agent proposes, another critiques, a third coordinates. It sounds like a good team.

Sometimes it is.

Research on test-time reasoning has found cases where multi-agent approaches outperform simpler strategies at comparable compute budgets. But recent work on agents acting for different users over shared resources shows the opposite pattern: adding agents can create stalling, conflicting actions and coordination overhead large enough to make the group worse than a single coordinator.

Those results are not contradictory. They describe different task structures.

More agents increase the amount of intelligence in the room and the amount of coordination the room requires.

Why people believe it

Human organizations often improve difficult work by adding specialization and peer review. It is natural to map that idea onto agents.

But software agents do not inherit effective organizational design automatically. They need protocols for authority, state, communication, conflict and termination.

What the evidence says

Multi-agent gains are most plausible when work can be cleanly decomposed, parallelized or independently checked. They become less reliable when agents compete over shared state, serve conflicting goals or repeatedly rewrite one another's work.

Anthropic's research on emerging multi-agent systems treats coordination itself as a major unresolved problem. The recent MAMUBench work shows that communication channels help but do not eliminate the gap.

The better question

Before adding an agent, ask: What unique work does this agent perform that cannot be done more reliably by a tool, a deterministic check or the existing agent?

If the answer is unclear, the extra agent may be architecture theatre.