跳至主要內容
返回技術博客

· 3 min read · ·

Cross-device AI agents: an engineering handoff checklist for GitHub AHP

GitHub introduced the engineering behind Agent Host Protocol on November 17. Here is how teams can evaluate session continuity through code identity, permissions, reconnection and verifiable delivery evidence.

  • AI Agent
  • GitHub Copilot
  • AHP
  • Delivery
On this page
  1. What changes when sessions continue across clients?
  2. Define evidence for a successful handoff
  3. Run three failure-oriented checks
  4. Start with a narrow pilot
  5. Common questions
  6. Sources and next steps

Starting an agent task on a laptop, checking it from a phone and returning to review it on a desktop changes what a handoff needs to contain. A conversation summary alone does not tell the next person which code is running or which action awaits approval.

On November 17, 2026, GitHub Engineering described Agent Host Protocol, an open protocol developed first in VS Code and adopted across GitHub surfaces. It lets clients interact with shared agent session state. The recommendations below are 8plus engineering analysis, not a benchmark or a promise that every third-party tool offers the same capabilities.

What changes when sessions continue across clients?

GitHub describes a server-authoritative state model in which clients submit actions and reconcile state after reconnecting. Separating session state from an editor window gives different interfaces a way to resume the same work.

For a delivery team, the practical question is whether the receiving person can make a correct decision. Conversation history is useful, but so are the active process, repository revision, pending changes and approval boundary.

Define evidence for a successful handoff

Consider an illustrative task: correcting date formatting in an order export. This is a test scenario, not a client case. A person switching from a laptop to a phone should be able to establish:

  • Task identity: the session, current objective and remaining steps.
  • Code identity: the repository, branch, base commit and uncommitted diff.
  • Execution state: whether a test is running, finished or disconnected.
  • Authority: who may inspect, edit, approve a merge or deploy.

These facts can appear in the interface or in accessible records. They should not depend on the original operator remembering what happened.

Run three failure-oriented checks

Disconnect and reconnect. Submit a task in a disposable environment, interrupt the client network and reconnect. Verify that replay does not create a duplicate ticket, message or deployment. Protocol-level reconnection does not establish that every external operation is idempotent.

Operate from two clients. Request a stop in one interface while submitting another instruction in the other. Check which state was accepted and whether its actor and timing are traceable. Use a resettable test task rather than payments or production deployment.

Change code during review. Produce a test result, then add another commit. Check that the earlier result is not presented as evidence for the new revision. Preserve the tested SHA, environment and execution time even when the entire conversation remains available.

Start with a narrow pilot

Choose reversible work with explicit completion conditions, such as a documentation change or a small reproducible bug. Require a diff, verification commands, actual results and unresolved questions. Keep human review before merging during the pilot.

Measure the time from taking over to making a decision, the number of repeated context questions and any duplicate side effects. These are proposed measures, not universal success thresholds. Establish your baseline before deciding whether a tool lowers handoff costs; generation speed alone misses the cost of reconstructing context.

Also rehearse exporting changes and taking over manually. An open protocol helps integration, but it does not establish identical feature coverage across products or unlimited background execution.

Common questions

Does AHP support mean an agent can deploy automatically?

No. Session continuity, execution permissions and deployment authorization are separate concerns. Production delivery still needs revision-specific evidence and a recovery path.

Should a small team switch tools immediately?

First establish whether cross-device handoffs are a real source of lost time. A team working from one workstation may benefit more from better requirements, tests and review evidence.

How does this relate to persistent agents?

Persistent agents concern how work continues; session continuity concerns how people inspect and take over that work. Read our Kiro bots discussion alongside this checklist when defining task boundaries.

Sources and next steps

SUBSCRIBE

新文章寄到你的信箱

每月約 2–4 篇,主題是 AI 導入、Web 工程與維運。只在有新文章時寄出,隨時可以退訂。

送出後會收到一封確認信,點擊確認才會開始寄送。確認信可能被歸到垃圾郵件,請留意一下。

用 RSS 訂閱

習慣用 Feedly、Inoreader 或其他閱讀器?加入這個網址即可。

https://www.8plus.app/feed.xml
開啟 RSS

想把文章裡的想法落地?

預約一場諮詢,讓我們針對你的系統與流程一起討論可行的做法。