# Let the next agent continue from checked work. Reviewed: 2026-10-18 Canonical: https://clankercloud.ai/learn/project-handoffs-between-ai-agents Hand off an agent task by saving its current sources, checked output, decisions, open questions, and operation status in the Project. Ask the next authorized agent to read that packet and confirm the next step before changing it. Clanker provides the shared work location; it does not automatically synchronize private conversations between assistants. ## What you are working toward A second task that can explain what is settled, what is uncertain, and what it may do next. ## Start with - The intended Project and current saved files - A concise handoff with source references and action status - A compatible, authorized receiving agent or native Workspace Agent ## Define the handoff as an artifact A handoff should answer five questions: what is the job, which sources are current, what has been verified, what remains unresolved, and what may happen next? Save those answers in a short file such as handoff.md. Refer to actual filenames and source dates; a statement that “everything is in the chat” gives a new agent no dependable starting point. Keep the packet focused on the next decision. A researcher handing off to a reviewer should provide an evidence table and a draft conclusion, not every search fragment. Include the key caveats that could change the result. The receiver needs enough context to challenge the work as well as continue it. ## Separate evidence, decisions, and suggestions Store source references beside consequential claims. Record whether a date is approved or proposed, whether an owner is named or missing, and whether a conclusion is a reviewer decision or an agent suggestion. These distinctions prevent the next assistant from turning a polished draft into an invented commitment. For a fictional exhibit-readiness task, a completed accessibility review can be a verified fact while a suggested opening date remains unresolved. The reviewer should see both without reading an entire private conversation. Use a source register for filenames and freshness, and a decision note for conclusions that someone actually accepted. ## Connect the receiver through its supported route A compatible Claude or ChatGPT client can read the packet through an authorized Workspace MCP connection. Another native Clanker Agent can work from the Project context with its selected tools. Confirm the receiver can read the actual files before asking it to edit them; public web access alone does not grant private Project access. The saved packet is the bridge between assistants. Private Claude and ChatGPT conversations are not automatically copied or merged. MCP permissions also operate across the connected user’s accessible Workspace, so a Project label in the handoff is a task boundary rather than a project-only OAuth grant. Sources: - [Clanker: external-agent access, permissions, and managed Sessions](https://clankercloud.ai/workspace-agents) ## Record completed and uncertain actions separately Include an action register when work involved writes, commands, or outside services. Name the known file, job, or task identifier and whether the result was inspected. Do not put credentials or private reasoning into the register. An action reported as started has a different recovery rule from one confirmed complete. If a save or command timed out, ask the receiver to inspect the original operation and destination before trying again. Use retained identifiers and the current interface’s revision and idempotency rules. “Continue the project” must not silently become “repeat every action the previous agent attempted.” Preserve uncertain status until evidence resolves it. ## Give one task ownership of the next change Two agents can read the same packet and produce conflicting edits if both believe they own the next version. Name the next deliverable and the task responsible for it. The receiving agent should read the current file and revision immediately before changing it, rather than write from an earlier memory of the handoff. For external MCP changes, follow the managed Session contract, record public progress, read back the saved result, and complete the task. If the observed revision is stale, reread and reconcile. A role division such as researcher then reviewer is a working convention; it is not evidence of automatic multi-agent orchestration. ## Test the handoff without the original conversation Open a fresh task and provide only the handoff location and intended job. Ask it to identify the current sources, reproduce one supported conclusion, name one unresolved issue, and propose the next permitted step. If it needs facts that exist only in the earlier chat, add the missing context to the packet. After the next task finishes, update the handoff to point to its checked result and new open questions. Keep earlier versions or a concise change note when they explain a decision. A useful chain of work remains understandable when an assistant changes, a runtime stops, or the person reviewing it was absent from the original conversation. ## Try this prompt Copy into a Workspace Session. Opening Workspace does not send it automatically. ```text In the selected Clanker Project, prepare a handoff for the next authorized agent. Read the current sources and saved output. Save handoff.md with the job, current filenames and versions, verified conclusions, unresolved questions, known operation identifiers and status, and one bounded next task. Exclude secrets and private reasoning. Inspect uncertain actions before suggesting retries. Read back the saved handoff. The receiving agent should confirm the packet before changing files; do not assume private chats synchronize. ``` ## Before using the result - The handoff points to current files that can actually be opened. - Verified facts, proposals, and unresolved questions are distinct. - Uncertain operations retain their original identifiers. - A fresh task can continue without the original private chat. ## Common questions ### Can Claude start a task and ChatGPT continue it in Clanker? A compatible authorized client can use deliberately saved Project context and outputs from the earlier task. Check the receiving client’s tools and permissions. Their private conversations do not synchronize automatically. ### Should I copy the whole conversation into a handoff? Usually a concise packet is more useful: current evidence, accepted decisions, saved results, open questions, and the next authorized action. Exclude secrets and private reasoning, and retain source references for material claims. ### What should the next agent do after a timeout? Inspect the original file, job, or operation using retained identifiers before repeating work. A timeout does not establish that a write or external action failed. Preserve uncertain status until it is reconciled. ## Sources - [Clanker: Workspace product behavior and feature availability](https://clankercloud.ai/faq) - [Clanker: external-agent access, permissions, and managed Sessions](https://clankercloud.ai/workspace-agents) - [Clanker: current agent instructions and HTTP access](https://workspace.clankercloud.ai/for-agents.md) ## Related - [A small packet the next agent can verify.](https://clankercloud.ai/templates/agent-handoff-pack) - [Design context that survives an assistant change.](https://clankercloud.ai/learn/design-agent-project-context) - [Give your agents a place to do lasting work.](https://clankercloud.ai/learn/ai-agent-workspace) - [Build an Agent you can test, reuse, and hand over.](https://clankercloud.ai/learn/ai-agent-factory) - [Give the Agent a test with an answer you can check.](https://clankercloud.ai/templates/agent-acceptance-test) [Open Workspace](https://workspace.clankercloud.ai) [Current pricing and usage](https://clankercloud.ai/pricing)