Skip to content
For your teamWhat one developer teaches an agent, a teammate’s agent can use.How it worksWhere a lesson starts and how it reaches every agent.Code reviewYour team’s knowledge, checked in every PR.SecurityWhat it sends, what it changes and how to leave.Book a call30 minutes with Roei, if you want a hand. Never required.
On your ownYour corrections follow you across agents and repos.Install itnpx harborloop init, in your repo.Every commandWhat the CLI does, and how to take it off again.
DocsPricingSign in
Connect your agents →
For teamsFor individualsHow it worksCode reviewSecurityDocsPricingSign inConnect your agents →
Home/Blog

Blog

  • Hand off between agents
  • Supermemory alternatives
  • PR review against AGENTS.md
  • Claude Code second brain
  • Why Copilot ignores it
  • Cursor Team Rules
  • Rules vs skills
  • Claude Tag vs Claude Code
  • GitHub Copilot Memory
  • Codex memories
  • Cursor memories
  • Claude Code auto memory
  • Claude Code team memory
  • Supermemory Company Brain alternatives
  • 3,905 files checked
  • Why Codex ignores it
  • Mem0 alternatives
  • Codex skills
  • Karpathy's LLM wiki
  • Auto mode
  • Memory or documentation?
  • Best company brain tools
  • Claude Code and AGENTS.md
  • Claude Code skills
  • Claude Code hooks
  • claude-mem
  • Claude Code best practices
  • Karpathy CLAUDE.md
  • AGENTS.md vs CLAUDE.md
  • Cursor rules
  • copilot-instructions.md
  • Why Claude ignores it
  • CLAUDE.md length
  • Committing CLAUDE.md
  • Agent memory compared
  • CLAUDE.md vs skills vs hooks
  • Make Claude Code remember
  • Review comments to rules
  • Slack decisions
  • Company brain
  • Context engineering, Claude 5
  • Context rot, stale rules
  • AGENTS.md research
  • ADRs for agents
  • Memory poisoning
  • Best MCP memory servers
  • Shared memory across agents
  • Growing CLAUDE.md
  • Which rules got used

Blog

How to hand off a task between Claude Code, Codex and Cursor

Published: November 13, 2026

To hand a task from Claude Code to Codex or Cursor, ask the agent you are leaving to write a short handoff note (the goal, what is done, what is next, what was decided), then start the next agent with “read the handoff and carry on”. No agent reads another's session, so the note is the bridge. Keep standing rules in AGENTS.md, where all three read them.

Why the next agent starts from nothing

Each agent keeps its sessions in its own place, in its own format, and none of them reads the others'. Resuming works only inside the same tool.

AgentWhere its sessions liveHow you resume, in that agent only
Claude Code~/.claude/projects/, one folder per projectclaude --continue or claude --resume
Codex CLI~/.codex/sessions/codex resume
CursorInside Cursor, per workspaceThe chat history in the Agent panel

What does cross over is what is in the repo. Codex and Cursor read AGENTS.md, and Claude Code reads it through a one-line @AGENTS.md import in CLAUDE.md (see AGENTS.md vs CLAUDE.md). That carries your rules. It does not carry the task you were halfway through.

The free way: a HANDOFF.md

Before you switch, tell the agent: “Write HANDOFF.md for the next agent: the goal, what is done, what is next, what we decided and why, and anything open.” Read it before you leave. The agent wrote it from memory, and that is where it goes wrong.

# Handoff

Goal: what the task is for, in one sentence.
Done: what works now, with the files touched.
Next: the very next step.
Decided: choices made, and the ones rejected with why.
Open: questions, failing tests, anything half done.
Git: branch, last commit, and whether it is pushed.

Then open the other agent in the same folder and say “read HANDOFF.md and carry on from Next”. Three habits make it hold:

  • Let git carry the code. Commit or at least stage before you switch. A note that says “done” next to a diff that says otherwise sends the next agent the wrong way.
  • Write down what you rejected. The next agent will otherwise try the approach you already ruled out.
  • Keep it out of the rules. A handoff is about one task and goes stale in a day. Add HANDOFF.md to .gitignore, and move anything that should hold for every task into AGENTS.md.

Should I compact or hand off?

Compacting keeps you in the same agent and the same session: it shrinks the conversation so you can keep going. A handoff is for when you leave the session, to another agent, a fresh session, or a teammate. If you are switching tools anyway, write the note first. A summary written while the agent still has the whole session in view is better than one written after compaction.

Handing the task to a teammate

The same note works, with one more step: push. A teammate cannot see files you have not committed or commits you have not pushed, so put the branch name in the note and push it, even as a work-in-progress commit. Pasting the note into the pull request description keeps it next to the code.

Tools that do the handoff for you

A few open-source tools read one agent's session and write the note or the prompt for the next. As their own listings describe them, on 12 November 2026:

ToolWhat it says it does
TStansel/handoffHands context from one agent to another, from the command line
OpenMOSS/claude-codex-handoffA file-based handoff protocol for Claude Code and Codex working in one project
crossbyA bridge across AI agents, including a handoff between them

Codex also has /import, which copies a Claude Code or Cursor setup across once (instructions and project memories). It moves your setup, not the task in flight (see Codex memories).

With Harbor: say handoff, then handoff pick

Harbor has handoffs built in for Claude Code, Codex, Cursor and GitHub Copilot. Type handoff where you stop. The agent writes the note in its reply, so you see it, and Harbor reads the branch, the commit, the uncommitted files and the commits not pushed from git itself. Open any agent in the same repo and type handoff pick, and it carries on from the next step. A handoff belongs to the repo, not the branch.

  • Forgot to save one? continue from claude (or codex, cursor, copilot) picks up that agent's last session in this repo on the same machine.
  • To a teammate: handoff @dana. Their agent fetches your branch, and asks first if they have uncommitted changes. If you have work they cannot see yet, the agent offers to push it.
  • How long: yours waits 72 hours, a teammate's 7 days, and every handoff is deleted after 90 days.

The full list of words is in the Handoffs docs. Harbor is shared memory for AI coding agents, and handoffs run on the hooks it already installs.

Questions

How do I hand off a task from Claude Code to Codex?

Ask Claude Code to write a short handoff note (the goal, what is done, what is next, what was decided and why), commit or stage your work, then open Codex in the same folder and tell it to read the note and carry on. Codex does not read Claude Code sessions, so the note is what crosses over.

Can Codex or Cursor read a Claude Code session?

No. Claude Code keeps sessions under ~/.claude/projects/ and resumes them with claude --continue or --resume; Codex keeps its own under ~/.codex/sessions/ and resumes with codex resume. Neither reads the other. Codex /import copies a Claude Code setup across once, not the task in progress.

What should a HANDOFF.md contain?

The goal in one sentence, what is done with the files touched, the very next step, the choices made and the ones rejected with why, open questions or failing tests, and the branch and last commit. Read it before you switch, and keep it out of git.

Should I use /compact or a handoff?

Compact to keep going in the same session. Hand off when you leave it, for another agent, a fresh session or a teammate, and write the note before compacting, while the agent still has the whole session in view.

Is there a tool that does the handoff automatically?

Open-source options include TStansel/handoff, OpenMOSS/claude-codex-handoff and crossby. Harbor has it built in for Claude Code, Codex, Cursor and GitHub Copilot: type handoff where you stop and handoff pick in the next agent, or handoff @name to give the task to a teammate.

Your team’s rules, in every agent, checked on every PR.

Harbor is shared memory for your team’s Claude Code, Cursor, Codex and Copilot. Free for you and one teammate. No card. No end date.

Start free →

Product

  • How it works
  • Every agent
  • Code review
  • The result
  • Security
  • For coding agents
  • Pricing

Developers

  • Docs
  • Quickstart
  • CLI
  • MCP server
  • Served and cited
  • Rule-file checker
  • Blog

Guides

  • Why Claude ignores CLAUDE.md
  • AGENTS.md vs CLAUDE.md
  • Does Claude Code read AGENTS.md?
  • Why Codex ignores AGENTS.md
  • Why Cursor ignores rules
  • Copilot instructions vs AGENTS.md
  • Shared memory across agents
  • Make Claude Code remember

Compare

  • vs Mem0
  • vs Zep
  • vs Unblocked
  • vs ByteRover
  • vs Cognee
  • vs Hindsight
  • vs OpenViking
  • vs Supermemory
  • vs Glen

Company

  • Questions
  • Contact
  • X
  • LinkedIn
  • Privacy Policy
  • Terms of Service
  • Data Processing Agreement
  • Refund Policy
© 2026 HarborConnect your agents