Agent handoff
By Mike Schreiber · Updated 31 October 2026
Agent handoff
Agent handoff is the transfer of responsibility and enough working context from one AI agent or session to another to continue a task safely. The useful unit is not necessarily a full transcript. It is a concise, current account of the objective, completed work, constraints, unresolved questions, evidence, and next action.
What a handoff carries
A practical handoff should identify the repository and branch, relevant files, decisions and their reasons, test results, blockers, and the receiving agent’s scope. Separate verified outcomes from claims and proposed changes. Include dates and evidence links so the recipient can check whether the state is still current.
Persistent instructions and task state serve different purposes. AGENTS.md or CLAUDE.md can describe standing conventions; a handoff explains where this particular task stopped. Codex discovers AGENTS.md through its instruction hierarchy. Claude Code’s direct support is version-dependent, requiring v2.1.277 or later, with loading also affected by configuration and nearby Claude instruction files. [1][2]
Transfer mechanisms and limits
Native session resumption preserves context within a supported product workflow. Claude Code can resume saved conversations, while /export produces readable plain text. Neither operation establishes a universal interchange format for another agent’s internal state. [3]
An MCP memory server offers another route: agents can retrieve recorded context through exposed tools or resources. Local versus hosted matters. A local server typically runs on the same machine over standard input/output; a hosted server runs remotely and commonly uses Streamable HTTP. Shared access depends on the server’s storage, configuration, and permissions, not merely on using MCP. MCP defines context exchange, not how applications manage memory or interpret retrieved information. [4]
Hark fits this workflow as a project record, not universal transcript transfer. Its briefs, memory, and handoffs read the same five blocks: Inventory, Health, Work, Authority, and Decisions. Records distinguish Agreed, Observed, and Proposed, with sources; acceptance does not independently verify an agent’s claim. [5]
Practical example
After investigating a payment retry bug, an agent leaves: “Branch fix/retries; duplicate-charge reproduction added; patch proposed but not merged; unit tests pass, integration tests not run; preserve the existing idempotency contract; next action: run the integration suite and review timeout behavior.” Link the patch, test output, and governing decision. The receiving agent checks the branch and evidence before editing, rather than treating the summary as proof.
Sources
[1] Codex: AGENTS.md