Agent Coordination

The coordination problem

A multi-agent architecture creates more surfaces where meaning can be lost: the user asks the lead agent, the lead frames subtasks, workers interpret those subtasks, tools return observations, workers compress results, and the lead synthesizes them. Each boundary is a chance to drop a constraint or invent a bridge that was never supported by evidence.

Coordination therefore has to be designed as a first-class product behavior. A useful design says what context is global, what context is private, what artifacts must be persisted, what messages can be lossy, and what decisions require a human or deterministic gate.

Shared memory and artifacts

Shared memory is not simply "put everything in a vector store." The useful unit is an artifact: a research note with citations, a test result, a plan, a patch, a data table, a policy decision, or a checklist. Artifacts preserve what the next agent needs without forcing it to replay every token of the prior trace.

This matters because subagents can otherwise become a game of telephone. Anthropic's production write-up recommends allowing subagents to write certain results to external systems and pass lightweight references back to the lead. That pattern preserves fidelity for structured outputs while reducing token copying through conversation history.

A practical memory design distinguishes durable task state from scratchpad reasoning. Durable state should be inspectable by humans and other agents: source URL, claim, timestamp, owner, confidence, and next action. Scratchpad text may help a model reason, but it should not be the only place a system stores commitments.

Message passing

Message passing is the protocol for moving work between agents. A weak message says "look into this." A strong message includes the task, the boundary, the source requirements, the expected artifact, the stop condition, and the questions that should be escalated instead of guessed.

Agent-to-agent messages should be shorter than full traces but richer than natural-language vibes. For research, a message might specify accepted source classes and citation format. For coding, it might specify files in scope, tests to run, and changes that are explicitly out of scope. For policy, it might specify which actions are blocked unless a human approves.

AutoGen is relevant because it treats multi-agent conversation as programmable: agents can be composed, customized, connected to tools, and arranged in interaction patterns. That flexibility is powerful, but it makes protocol design more important, not less.

Handoffs

A handoff transfers control. In OpenAI's terminology, multi-agent systems include manager patterns where a central agent calls specialized agents as tools, and decentralized patterns where agents hand workflow execution to one another. The distinction matters because control determines who sees the user, who owns the final answer, and where guardrails run.

Handoffs are appropriate when the next agent has a different operating policy or domain. A refund agent, a legal-review agent, and a technical-troubleshooting agent should not be the same role if their tools, risk thresholds, and escalation paths differ. But every handoff should carry enough state for the receiving agent to act without rediscovering the world.

Observability and evaluation

Multi-agent traces need observability at two levels. First, you need normal operational telemetry: tool errors, retries, latency, cost, and stalled tasks. Second, you need behavioral telemetry: which agent delegated what, what sources were selected, what artifacts were accepted, and which paths correlate with bad outcomes.

Anthropic's write-up is clear that traditional evals are harder because valid agent paths differ between runs. The answer is not to give up on evals. It is to evaluate final outcomes, citation accuracy, source quality, process reasonableness, and tool efficiency. Human testing still matters because human reviewers catch source-selection bias and edge cases that automated judges miss.

The coordination standard is therefore concrete: every important action should be traceable to an instruction, observation, artifact, or explicit assumption. If you cannot reconstruct why an agent did something, you cannot reliably improve the system after it fails.