Agent Handoffs

In a multi-agent system, agents rarely operate in complete isolation. At some point, one agent's work becomes another agent's starting point. This transition — when control and context move from one agent to another — is called a handoff. Handoffs seem simple in theory but are one of the most error-prone moments in any orchestration system. A handoff that loses context produces a downstream agent that starts from an incomplete picture. A handoff that transfers too much irrelevant information overloads the receiving agent's context window. Getting handoffs right is a core engineering discipline in multi-agent design.

What a Handoff Transfers

A handoff is not just passing a file or a text string. It is a carefully structured transfer of three distinct things: the goal (what the receiving agent must accomplish), the state (what has been learned or produced so far that the receiving agent needs to know), and the constraints (time limits, formatting requirements, scope boundaries, or other rules the receiving agent must respect). Consider a customer service pipeline. Agent A is a triage agent that reads an incoming support ticket, classifies the issue type, and extracts key facts. When it hands off to Agent B, a specialist resolution agent, the handoff must include: the classification result, the extracted facts, the customer's account history if relevant, and any constraints the resolution agent must respect (e.g., only offer refunds up to a certain value). If the handoff omits the extracted facts, Agent B must re-read the raw ticket — wasting context and time, and risking different conclusions. If it omits constraints, Agent B might offer a resolution that violates policy.

Handoff Formats: Structured vs. Natural Language

Handoffs can use natural language summaries or structured data formats — JSON, YAML, XML, or domain-specific schemas. Each has tradeoffs. Natural language summaries are easy for the receiving agent to incorporate into its reasoning context. They read like briefings from a colleague. But they are ambiguous, hard to parse programmatically, and can silently omit critical information if the sending agent's summarization is poor. Structured formats are precise, machine-readable, and easy to validate with a schema. A receiving agent (or the orchestration layer itself) can verify that required fields are present before the agent starts work. Structured formats also enable logging and auditing — you can inspect every handoff message in a system trace and know exactly what information each agent received. The tradeoff is that LLM-based agents must be explicitly instructed to produce and consume structured formats, and the system prompt overhead increases. Best practice in production systems is structured handoffs with a natural language 'summary' field: the machine parses the structured fields for routing and validation, and the receiving agent reads the summary field for readable context.

Failure Modes in Handoffs

The most common handoff failure is context loss: information that the sending agent has but considers obvious is omitted from the handoff, and the receiving agent proceeds on an incomplete picture. This is analogous to a surgeon handing off to a recovery nurse without communicating which medications the patient received during the procedure. The second failure mode is context contamination: the receiving agent is given too much irrelevant history, and its reasoning drifts toward that history rather than the current goal. This is especially pernicious in LLM-based agents, where long context windows can cause the model to anchor on early information even when later information should take precedence. The third failure mode is format mismatch: the sending agent produces output in a format that the receiving agent or the orchestration layer cannot parse. This causes the downstream agent to fail immediately, often in ways that are difficult to debug if the orchestration system does not log the raw handoff payload. A fourth failure mode is goal drift: the handoff correctly transfers state but describes the goal ambiguously, and the receiving agent pursues a subtly different goal than intended. Explicit, unambiguous goal statements — preferably phrased as acceptance criteria rather than open-ended instructions — are the mitigation.