When Do Agents Work Well Together?

Agents work well together under three conditions: each agent can see enough about what other agents are doing, the system prevents collisions, and all agents are accountable to one final objective even when intermediate goals diverge. If any of those conditions is weak, multi-agent setups often become expensive theater.

This boundary is easier to understand through Coase's theory of the firm. Firms exist when internal coordination is cheaper than repeated market transactions. Multi-agent systems face the same tradeoff. A single agent loop is cheap to start and simple to supervise. A team raises setup cost, but can lower recurring cost if it reduces context rebuild, ownership confusion, and avoidable review cycles.

An operations research lens turns this into a concrete decision problem. You are optimizing throughput and reliability under constraints: finite processing capacity, task dependencies, and variable rework rates. Concepts from queueing theory, Little's law, and critical path method are useful because they force one practical question: does decomposition shorten the real bottleneck, or only add more waiting and handoff loss. Teams win when specialization reduces cycle time and tail risk faster than coordination overhead grows.

A compact rule is:

Use a team when expected single-agent cycle time + rework + failure cost is greater than expected team cycle time + rework + failure cost + setup cost amortized over repeated runs.

That rule only works if shared context and collision control are built into execution, not left to convention. Shared context means each role can access current plan state, artifact status, and dependency state before acting. Collision control means explicit mechanisms such as locks, leases, idempotency keys, write scopes, and commit gates.

Game theory explains why this discipline matters. In a multi-agent workflow, each role is a player with partial information and local incentives. Without mechanism design, locally rational behavior can still degrade global outcomes. This is the core point of mechanism design: the rules define the equilibrium. If local payoff is tied to speed alone, quality collapses downstream. If payoff is tied to terminal outcome quality, coordination becomes stable.

The highest leverage mechanism is authority asymmetry. Proposal rights should be broad, but commit rights should be narrow. Many agents can propose; one accountable path authorizes high impact side effects. That single constraint limits diffusion of responsibility and keeps failure surfaces legible.

Recent research supports this view. AgentBench shows persistent weakness on long horizon agentic tasks. MultiAgentBench shows protocol design materially changes outcomes. Why Do Multi-Agent LLM Systems Fail? identifies recurring coordination and system design failures across many traces. Single-agent or Multi-agent Systems? Why Not Both? shows hybrid routing can outperform both always-single and always-multi policies under realistic cost constraints.

So when do agents work well together? When coordination is treated as a first-class systems problem: enough shared context to act coherently, hard controls to prevent stepping on each other, and one measurable goal that dominates local preferences.