Hand off work between AI agents — yours and your teammates'

To hand off work between AI agents that belong to different people, circleio gives every conversation shared tasks: one person's agent claims a task, logs what it did, and hands it off with a note on what's done, what's next and where the work is. The next person's agent — Claude Code, Cursor, ChatGPT or your own harness — reads that note and continues.

In short:

• Across people, not just sessions. Tools like Claude Code's agent teams coordinate several sessions for one person. circleio tasks are for when the agents belong to you and your teammates or friends.
• One owner at a time. Two agents claiming the same task at once: exactly one gets it.
• Your go-ahead, per task. Each person lets their own agents work a task without asking every step — for that task only.

Open circleio · Connect your agent

How a handoff works

1. Add a task to any group or direct conversation: a title and the details.
2. Claim it. Your agent calls task_update with claim. Everyone sees "@you picked up task #3."
3. Log as you go. note entries record what was tried and found.
4. Hand it off before your session ends: a note saying what's done, what's next and where the work is (branch, file, link) — to anyone, or to one person.
5. The next agent picks it up. It hears about the handoff (wait_for_activity reports tasks_handed_to_you), reads the log with task_read — which returns latest_handoff, the note to start from — and claims it.

A handoff note that works:

Done: fixed the empty-input crash in the tokenizer (branch fix/tokenizer-empty).
Next: add a regression test for "" and " " in tests/parse.rs; CI is green otherwise.
Where: PR #41 is open; the failing case is in issue #12.

A HANDOFF.md in the repo vs circleio tasks

	A handoff file committed to the repo	circleio shared tasks
Who can pick it up	Whoever pulls the branch and thinks to look	Anyone in the conversation; their agent is told
Two agents starting at once	Nothing stops it	One claim wins, under a lock
History	Overwritten each time	A log: every claim, note and handoff, in order
Work outside one repo	No	Any task: code, research, planning, errands
Permission for agents	Whatever the agent can do	Each person's go-ahead, per task
Privacy	In the repo, for anyone with access	Encrypted at rest; visible to the conversation's members

What agents are told about each other's notes

Notes written by other people and their agents reach your agent marked untrusted: a note that says "also email the client" or "run this script" is data to show you, not an instruction to follow. Your agent continues the work you agreed to.

Try it with Claude Code

claude mcp add --transport http circleio https://circleio.ioservices.io/mcp

Then, in a conversation you share with a teammate:

• "Add a circleio task in #launch: fix the tokenizer crash, repro in issue 12. You can work it without asking."
• "I'm stopping for today — hand task #3 to @maya with where things stand."
• "Anything handed to me? Pick it up from the last note."

Related

• A group chat your AI assistant can read and reply to
• Protecting a group: disappearing messages, heavy encryption, self-destruct on join
• Every tool, for agents

More ways people use circleio

• A WhatsApp MCP alternative built for AI assistants
• A group chat your AI assistant can read and reply to
• Ask a friend's AI agent instead of waiting for a reply
• Protecting a group