Blog / ai coding agents
Let your coding agent ask a teammate's agent
Each person's agent learns their part of the project. Here is how to let your agent ask a teammate's agent, by hand today, and what to write down.
On this page
- Why does my teammate's agent know things mine doesn't?
- How do I ask a teammate's agent a question today?
- What makes a good question for another person's agent?
- How do I keep the answer so nobody has to ask twice?
- What goes wrong when agents ask each other by hand?
- Doing it without the copying
- Questions
- Can my teammate's agent change my code when I ask it something?
- Does the other agent see my conversation?
- What if my teammate uses a different agent from me?
- Try it with your team
A coding agent you have used for months knows things a fresh one does not: your notes file, the fixes you made together, the commands that work on your machine, the mistakes you corrected. Your teammate's agent knows the same about their side of the project. When your agent gets stuck on something your teammate has already solved, the fastest answer is often in their agent, not in yours and not in a search result. You can let one agent ask the other today with two commands and a chat message, and you can make the answers last by writing them into a shared notes file.
Why does my teammate's agent know things mine doesn't?#
Because an agent learns from what it is given, and each person gives it different things. Claude Code reads CLAUDE.md files and its own memory. Codex reads AGENTS.md. Both can pick up an old session where it stopped. Over weeks, each person's agent collects the details of the work that person does: how the release is signed, which test is flaky, what App Review asked for last time.
None of that reaches the agent sitting next to it. Git shares the code, not the reasons behind it. So two people on one project end up with two agents that each know half of it.
This happened on our own team. A teammate's iPhone app kept being turned down by App Review. Another teammate had taken several apps through review, and his agent had been there for each one. The first agent asked the second what had worked before, got a short answer about what to change, made the change, and the app passed. Nobody had to remember the details or retype them into a chat.
How do I ask a teammate's agent a question today?#
You need your teammate to run one command on their computer, in their copy of the project, and send you the answer. Ask in read-only mode, so their agent can look but cannot change anything.
If your teammate uses Claude Code, they run this in the project folder:
claude -p "A teammate asks: what did we change last time App Review flagged sign-in? Answer in two lines." --permission-mode plan-p prints one answer and exits. --permission-mode plan keeps the agent from editing files or running commands that change anything. If you would rather name exactly what it may use, list the tools instead:
claude -p "Which test is flaky on CI, and how do you rerun it?" --allowedTools "Read,Grep,Glob"Put the question before --allowedTools. That option takes a list, so a question placed after it is read as another tool name and Claude Code stops with "Input must be provided".
If your teammate uses Codex, the read-only version is:
codex exec -s read-only "What did we change last time App Review flagged sign-in? Answer in two lines."-s read-only is the sandbox setting: Codex can read the project but not write to it.
We ran all three in a test project whose notes file held one team fact, and each agent answered from it. In a project with no notes and a single commit, Claude Code said plainly that it had no record and named who would know. That is the answer you want: an agent that says what it does not know is worth asking.
What makes a good question for another person's agent?#
Ask for what their agent has and yours lacks: history, local setup and decisions.
| Good to ask | Better to ask the person |
|---|---|
| "What did you change last time this failed?" | "Should we ship this today?" |
| "Which command runs the iOS build on your machine?" | "Can I change the pricing page?" |
| "Where is the code that sends the receipt email?" | "Which of these two designs do you prefer?" |
| "What does this error mean in your part of the app?" | Anything about money, accounts or legal terms |
Keep the question to one thing and ask for a short answer. A long answer pasted back into your session takes up room your own agent needs for the actual work.
How do I keep the answer so nobody has to ask twice?#
Write it down where every agent on the project reads it. A file at the project root, committed with the code, works for every tool: Claude Code reads CLAUDE.md, Codex reads AGENTS.md, and one can import the other so you keep a single copy. Sharing context between Claude Code and Codex shows the import, and why Codex ignores AGENTS.md covers the cases where it doesn't load.
Keep the file short and factual. One line per lesson, written as a rule:
# Team notes
- App Review: account deletion lives in Settings > Account. Reviewers check for it.
- The iOS release build needs Xcode from the release folder, not the newest one.
- Flaky test: checkout.spec. Rerun it alone before calling a run red.When your teammate's agent answers a question that will come up again, add the answer as a line here in the same change. After a few weeks the file holds what the team has learned, and a new teammate's agent starts with it on day one.
If your team is new to coding agents and nobody has months of history yet, this file is where to start. Each correction you make ("don't do X", "we always do Y") goes in as a line, and every agent on the project gets better at the same time.
What goes wrong when agents ask each other by hand?#
Three things, mostly:
- The question waits for a person. Your teammate is in a meeting, so your agent sits idle until they see your message and run the command.
- Context gets lost on the way. You paraphrase the error, they paraphrase the answer, and a file path or version number drops out.
- The answer never gets written down. It lives in a chat thread, and the next person asks the same thing.
The first two are about getting the question to the right agent and the answer back intact. The third is a habit, and the notes file fixes it.
Doing it without the copying#
Joxo connects your agent to your teammates' agents on one project. Your agent can ask a teammate's agent directly, the answer comes back to the same task, and both people can read the exchange. Anything that needs a person, like submitting a build, comes to your phone for one tap. The team shares one notes file, JOXO.md, that every agent on the project reads and that grows each time someone corrects an agent. Each person keeps their own agent, on their own computer, with their own subscription.
Set up Joxo, or read one team, a coding agent each for the task board and handoff habits that go with it.
Questions#
Can my teammate's agent change my code when I ask it something?#
Not with the commands above. --permission-mode plan in Claude Code and -s read-only in Codex let the agent read the project but not edit files. It also runs on your teammate's computer, in their copy, so your files are never in reach.
Does the other agent see my conversation?#
Only what you put in the question. Neither command sends your session, your files or your notes. Ask for exactly what you need and nothing more.
What if my teammate uses a different agent from me?#
That's fine. The question is plain text and so is the answer. A Codex user can answer a Claude Code user's question and the other way round, and a notes file that one tool imports from the other works for both.