Use a Ready for Agent status to hand Notion tasks to Claude (2026)

The trick: make a status column the handoff point

A product task can now move from a Notion board into a coding agent without a Zapier or n8n relay. Notion's July 1 release added External Agents to the shared workspace surface; Claude and Cursor were the first supported agents. 1

Use one status value as the boundary:

Notion task → Status = Ready for Agent → Claude session → GitHub pull request → human review

The status change starts the run. Claude posts progress on the task, works in the permitted repository, and returns a pull request link when the task calls for a code change. Notion's task-board guide also supports an @Claude mention when a human needs to add context or start work outside the normal status flow. 2

Before you build it

Have these pieces ready:

A Business or Enterprise Notion workspace. Claude agents are currently in beta. A workspace owner must enable them for Enterprise and HIPAA workspaces, and the rollout may still be gradual. Open Settings → Notion AI → Agent → Manage external agents if Claude does not appear. 3
Notion credits. Claude runs consume Notion credits based on the amount of reading, tool use, and work in each run. You do not need an Anthropic account or an Anthropic API key. 3
A GitHub connection. The coding-task-board setup asks for a GitHub personal access token with read and write permissions for both Contents and Pull Requests. Choose the repositories the agent may access. Without the pull-request permission, the agent can read code but cannot open a PR. 2
A task database with a Status or Stage property. Create statuses such as Backlog, Ready for Agent, In Progress, In Review, and Done. Notion's guide says the trigger works with any database that has a status or stage property, so you can add it to an existing roadmap or bug database. 2

Keep the task row as the control surface. Give it the brief, acceptance criteria, repository name, and the links to any shared specs. Do not make the agent search the whole workspace for context that should have been attached to the task.

What actually fires

Part	Configuration	Result
Trigger	The task's `Status` or `Stage` changes to `Ready for Agent`. 2	Claude starts an agent session for that task. 2
Action	Claude reads the shared task and the permitted GitHub repository. 2	The agent posts progress and can plan, edit, test, and prepare the change. 2
Write-back	The agent comments on the task and opens a PR when the work is a code change. 2	The task becomes the audit trail; GitHub holds the diff and checks.
API boundary	No Notion REST call or webhook is needed for this setup.	The status trigger and GitHub connection are configured in Notion; the PAT supplies the `Contents` and `Pull Requests` permissions. 2

The last row is the important implementation distinction. This pattern uses Notion's built-in agent trigger, not a hand-written PATCH /v1/pages loop. If you want a raw API-driven webhook pipeline, use a different pattern. Here, the stitching happens at the status change and the connected GitHub action.

Configure the handoff

Open Agents in the Notion sidebar and click New Agent → Claude. Choose the Coding task board starting point. The template supplies the task-board shape and its initial trigger; you can also start from a blank agent if your database already exists. 2
Open Tools and access, connect GitHub, paste the PAT, and select the repositories Claude may use. Keep the repository list as narrow as the task type allows. 2
Check the trigger for the Ready for Agent value. If you use an existing database, map the trigger to its Status or Stage property instead of creating a second task board.
Share the task board and the source pages with the agent. Give source specs the lowest access level that lets Claude read and interact with them. Give the task database edit access only if Claude must update status, comments, or links. Agent permissions are set per agent and do not come from the person who starts the run. 3
Replace the template instructions with rules that make missing context stop the run rather than invite a guess. For a product team, this is a useful starting block:

You are the engineering handoff agent for this task board.

Use only the task, its attached Notion pages, and the GitHub repositories shared with you.

Before changing code:
1. Confirm that the task has a problem statement, acceptance criteria, target repository, and test expectation.
2. If any required field is missing, leave a comment beginning "Needs input:" and stop. Do not open a pull request.
3. If the fields are complete, set Status to "In Progress" and state the plan in a task comment.

During implementation:
4. Inspect the relevant code and existing tests before editing.
5. Keep the change inside the task's acceptance criteria. Ask a question in the task if the scope is ambiguous.
6. Run the repository's relevant tests and report the command and result.

When the work is ready:
7. Open one pull request against the configured base branch. Include the task link, changed files, test result, and any follow-up risk.
8. Add the pull request link to the Notion task and set Status to "In Review".
9. Never merge the pull request or mark the task Done. A human owns review and merge.

The instruction block creates a second gate after the status trigger: Claude can start automatically, but incomplete product work stops before it reaches GitHub.

Test it with one small card

Create a low-risk bug or documentation change with the repository, acceptance criteria, and test expectation filled in.
Move the card to Ready for Agent. Confirm that an active Claude session and progress comments appear on the task. The guide says teammates can open the live session, see its transcript and tool calls, and add context while the run is active. 2
Run the failure test: remove the repository or acceptance criteria, then move a second card to Ready for Agent. The expected result is a Needs input comment and no PR.
Let the complete card run. Verify that the PR points back to the Notion task, the test result is present, and the task moves to In Review only if your instruction produced that write-back.
Review the diff, checks, preview, and scope as you would review a teammate's PR. The agent can prepare the change; it does not replace the merge decision. 2

Gotchas

The feature may be enabled but still invisible. Claude agents are in beta and rolling out gradually. Check workspace controls and onboarding before changing the database or rebuilding the workflow. 3

A task mention is not a permission grant. The agent can see only the pages, databases, and connections shared with it. A page name in the instruction block does not grant access. 3

Credits follow the run, not the number of cards. A batch of large tasks can consume more credits because Claude reads more context and uses more tools. Start with one card, inspect the run, then decide whether a batch trigger fits your team. 3

Claude cannot browse the web in a session. Put the source material in shared Notion pages or the connected repository before the status change. Do not make a live research step part of a coding handoff that has no web access. 3

The PAT is an execution boundary. Read access lets Claude inspect code; Pull Requests write access lets it create the review artifact. Keep the repository allowlist and token permissions narrow, and leave merge rights with humans. 2

Once the first card passes the missing-context test, the workflow has a clean handoff: PM work becomes a status change, engineering work becomes a visible agent session, and review starts from a linked pull request rather than a second intake queue.

참고 출처

1
Notion 3.6: External Agents, HTML blocks, and morenotion.com
2
How to set up Claude agents on your team's Notion task boardnotion.com
3
Use Claude agents in Notion (beta)notion.com