# A useful agent handoff includes the work Make agent handoffs reviewable with stable artifact references, revision identity, checks, limitations, and explicit acceptance. agentlist.iocollaborationPublished Oct 22, 20264 min read A message saying “finished” leaves the next agent with several questions. Where is the result? Which revision was checked? What remains uncertain? Can the recipient inspect the work, or must it trust the summary? The answers determine how much effort delegation saves. If the recipient has to reconstruct the task and locate its output, part of the work has simply moved into the handoff. ## Pass an artifact the recipient can inspect [Anthropic's account of its multi-agent research system](https://www.anthropic.com/research/multi-agent-research-system) describes workers storing artifacts and returning references. The authors present this as a way to reduce information loss and repeated copying through the coordinator. It is an account of their system, not proof that every task benefits from more agents. For a practical handoff, the reference must resolve for the recipient. A local path on another machine is insufficient unless the receiving workflow can retrieve it. A document link without access creates a similar dead end. Before handing over work, check how the recipient will open it. Prefer a stable artifact ID or a revision-specific reference when later edits could change what the reviewer sees. ## Make the result and its status explicit Consider a coding task. A useful handoff might contain: ``` Task: Reject empty CSV headers with a clear error. Artifact: Reviewable branch or patch reference. Revision: Exact commit ID checked by the worker. Changed scope: Parser and parser tests. Evidence: Relevant test command, result, and log reference. Limitations: Import UI behavior has not been checked. Next action: Review the diff and test the import UI. Acceptance: Pending reviewer inspection. ``` This is a proposed template. Fill it with actual evidence rather than copying the example's success language. If checks did not run, say why. If the worker returned only a partial result, preserve that status. For research, substitute the report, source record, and unresolved claims. For a spreadsheet, include the workbook revision, changed sheets, and checks on formulas or totals. The fields should help the recipient decide what to inspect next. ## Keep acceptance separate from delivery The sender knows that it delivered an artifact. The recipient determines whether that artifact satisfies the assignment. Combining those states can hide incomplete work behind a successful message exchange. Have the recipient open the artifact and test the important claim. A parser test log should match the supplied revision. A research conclusion should be traceable to the source cited for it. A saved chart should reflect the supplied dataset. When the recipient finds a problem, send a bounded correction that identifies the failed criterion. “The patch changes the parser, but the empty-header fixture still passes through” is more useful than “please improve this.” ## Use summaries to direct attention The summary still matters. It should explain the result and point out the parts that deserve review. It should not become the only surviving copy of the work. Start with one handoff in a real workflow. Count the questions the recipient must ask before reviewing the result. Improve the missing fields or access path. If the handoff costs more than the task, consider keeping the work with one agent until the task boundary is clearer. For a coding handoff, continue with [Ship a small change with a coding agent](https://www.agentlist.io/learn/guides/ship-with-a-coding-agent). The [Devin vs Windsurf workflow comparison](https://www.agentlist.io/articles/devin-vs-windsurf) also explains what to inspect when work returns from a separate environment. ## Give the next reviewer a clear brief. Use the fieldbook’s task-brief and evaluation worksheets to record the assignment and assess the returned work. [Get the brief and evaluation worksheets](https://www.agentlist.io/ebook?from=agent-handoff-evidence) --- Source: [A useful agent handoff includes the work | agentlist.io](https://www.agentlist.io/articles/agent-handoff-evidence). This is the public page rendered as Markdown; interactive controls require the website.