Coordination patterns

Designing an autonomous agent system means deciding which unit of decomposition fits each piece of work. The same operation, "check the weather", can reasonably be modeled as a tool, a task, or its own agent, and the choice has real consequences for context, persistence, and reuse. This section explains what each abstraction is for, then offers a rule of thumb for choosing between them.

Control transfers between agents. One agent is active at a time. Context accumulates or transforms as it moves through the chain.

Behavior: This is the simplest multi-agent pattern, and the most coherent. A single thread of reasoning runs throughout. The trade-off is path dependency. Early decisions constrain later ones, and errors compound rather than correct. The sequence order matters. Research-then-plan produces different results than plan-then-research. Real workflows are often graphs rather than linear chains, which is where sequential composes with delegative.

When to use: Workflow processes with clear stages and specialization at each stage. Better for refinement, where each stage improves on the last, than exploration. The linear sequence is easy to trace and interpret.

Example: A triage agent receives a customer request, determines the category, and hands off control to the appropriate specialist agent. The specialist now owns the interaction.

A coordinator assigns subtasks to workers. Workers operate in isolated contexts. Results flow back for synthesis.

Behavior: Context isolation is the defining feature. Each worker gets a focused context and can go deep on its subproblem without distraction or influence from the others. But isolation also means no cross-pollination. The responsibility for coherence falls entirely on the coordinator. It needs to perform well at both decomposition and synthesis. A variant is competitive delegation, where the coordinator assigns the same task to multiple workers and the best result is selected. At large scale, with many agents and statistical selection, this blurs into the emergent pattern.

When to use: Tasks that decompose into distinct subtasks benefiting from isolated, focused contexts, or when independent perspectives are needed. Good for parallel execution to reduce latency, or broad exploration with many workers generating independent attempts.

Example: A research coordinator delegates fact-gathering to a researcher and trend analysis to an analyst. Both work in parallel with isolated contexts. The coordinator synthesizes their findings into a brief.

Peer agents share context and communicate directly. They work together on a common problem.

Behavior: The strength of this pattern is mutual awareness. Agents can build on, question, and correct each other. Debate can surface better answers than any single agent would find. But shared context also enables groupthink: agents can converge on ideas too early, reinforce each other’s biases, or defer to whichever agent communicates first or most confidently. The more interdependent the collaboration, the more it resembles a single agent rather than multiple agents working in parallel.

When to use: Interdependent subtasks where agents need to see each other’s work, or when quality benefits from debate or review between peers. Good for error-catching through challenge, or when different expertise needs to be actively integrated, not just combined afterward.

Example: A team lead decomposes a project into tasks. Developer agents claim tasks from a shared list, work on them independently, and message peers when coordination is needed. The lead monitors progress and disbands the team when done.

Moderation as a controlled variant. Moderated turn-taking is a more structured form of the collaborative pattern. Members still share context, but a moderator controls who speaks when, sees the full transcript, and decides when the conversation ends. This trades some of the autonomy of a self-coordinating team for a predictable cadence and a single point that drives the discussion. Use it when the value comes from a shaped exchange (peer reviews, negotiations, panel discussions) rather than from members claiming and completing work in parallel.

Many agents operate in parallel with minimal individual context, following simple rules. Each agent modifies a shared environment, and that environment influences what other agents do. Complex behavior emerges from these indirect interactions.

Behavior: This pattern is for scale and exploration, and is the most difficult to implement effectively. The system’s behavior is statistical, the aggregate of many simple actions. That makes it resilient, with no single point of failure, but also unpredictable, since behavior emerges from interactions that were not explicitly designed. Selection turns a high volume of independent contributions into a useful result. Emergent selection lets better contributions get reinforced as other agents build on them. External selection introduces something outside the swarm to evaluate and curate the output. In practice, you may want both.

When to use: Large-scale problems with many similar agents and tolerance for probabilistic outcomes. Good for broad exploration, resilience to individual failure, and avoiding early convergence on a solution.

Example: A brainstorm team generates ideas on a shared board. Each agent contributes independently. A lead curates the results, with the final output emerging from accumulation and refinement rather than explicit coordination.

Context management is the primary motivation for multi-agent systems. An agent’s context shapes its behavior. What it sees determines what it attends to, how it reasons, and what solutions it considers.

As tasks grow more complex, a single agent’s context becomes diluted. Multi-agent patterns address this by scoping context so that each agent operates with clarity. The coordination pattern you choose is a context management strategy.

A well-designed single agent with appropriate tools can accomplish a lot. Multi-agent patterns introduce real overhead in tokens, in coordination, and in complexity.

Consider multiple agents when:

Multi-agent systems can also enable more dynamic behavior. Each execution is shaped by the interplay between agents rather than a single initial prompt. A delegative coordinator can tailor its decomposition per execution. Collaborative agents adapt to each other’s contributions as they work. Emergent agents respond to evolving shared state. The tradeoff is predictability. The more dynamic the coordination, the harder it is to anticipate and test the system’s behavior for any given input.

Start with the simplest approach that works, and add multi-agent complexity when there is evidence that a single agent is hitting its limits. It can also be worth experimenting with multi-agent approaches early to understand the tradeoffs, particularly when explainability or auditability are requirements.

The coordination patterns compose at different levels. Some examples:

Sequential pipeline with delegative stages. A document flows through stages: research, analysis, writing, review. Each stage can internally delegate to multiple workers. The analysis stage fans out to workers examining different aspects in parallel, then synthesizes before passing forward.

Delegative coordinator with collaborative teams. A coordinator delegates to specialist teams. Within each team, agents collaborate, debating approaches, catching errors, building on each other’s work. The coordinator only sees the final output from each team.

Collaborative debate with delegative research. Agents debate an approach, but when they need evidence they delegate fact-gathering to workers. Workers return findings and the debate resumes with new information. Useful when a decision process needs to be informed by research that can happen in parallel.

Delegative coordinator with sequential workers. A coordinator fans out to workers, but each worker internally runs a multi-step pipeline (research, draft, self-review) before returning results.