Blog / ai coding agents
AI agent collaboration: coordinate coding agents in a team
Give every agent the same rules file, one task board where each task has one owner, a decision log and handoff notes. Agents share facts; people decide.
On this page
- Why do AI coding agents on one team duplicate work?
- What should a team's coding agents share?
- How do I set up one task board for different agents?
- Should AI agents ask each other, or ask a person?
- How do I keep agents from waiting on a person?
- Who reviews pull requests from several agents?
- When this is not enough
- Frequently asked questions
- How do teams coordinate several developers using different AI coding agents?
- Should AI coding agents talk to each other directly?
- Do all teammates need to use the same coding agent?
- What should AI coding agents on a team share?
- Try it with your team
To coordinate AI coding agents across a team, give every agent the same rules in one instruction file, keep one task board where each task has a single owner, log decisions where every agent reads them, and pass work on with a handoff note that names the branch and commit. Let agents exchange facts, but send choices to the person who owns that part of the product. Keep questions small enough to answer from a phone, so no agent sits idle while its person is at lunch.
Picture three people building one product over a weekend (an illustration, not a customer). Priya uses Claude Code, Sam uses Codex, Jo uses Cursor. Every agent is good, and every agent knows exactly one person's side of the story.
By Saturday afternoon, Priya's agent has built a token refresh flow. Sam's agent has built a different one, because Sam's choice never reached anywhere Priya's agent could read. Jo's agent has named the same table two ways in two migrations. Each agent did what it was told with the information it had.
Why do AI coding agents on one team duplicate work?#
Because each agent only knows its own window. Git syncs the code, eventually, but not the reasoning behind it, and not until somebody pushes. The gap between "I decided this" and "everyone's agent knows it" is where a weekend gets lost.
Most advice for several agents on one repo is written for one developer: give each agent a worktree, name file owners, review each branch. That still applies, and running Claude Code sessions in parallel covers it. A team adds three harder problems on top. The agents can't talk to each other. Their people are busy and not always at a keyboard. And the agents may be different tools that read different instruction files.
What should a team's coding agents share?#
Less than you would guess. Not transcripts: they are long, private and mostly irrelevant to anyone else. Four short things carry nearly all the value:
| What | Example | Who needs it |
|---|---|---|
| Decisions | "Sessions are cookies, not JWTs." | Every agent, before it writes code |
| Ownership | "Sam's agent has search pagination." | Any agent choosing its next task |
| Blockers | "Export waits on the refund question." | The person who can unblock it |
| Handoffs | "feat/csv-export at a41c9e2; next, fix quoting." | Whoever picks the work up |
They are also what a human teammate needs, so writing them down helps even before any agent reads them.
How do I set up one task board for different agents?#
A shared list works to start: a GitHub Project, an issue tracker, a TASKS.md. The tool matters less than a few rules, and the rules go in the instruction file every agent reads. Put them in AGENTS.md and import it from CLAUDE.md, so Codex and Claude Code follow the same text; sharing context between Claude Code and Codex explains the import. For any other agent, check which file it reads.
## Team workflow
- Before starting work, read TASKS.md and DECISIONS.md.
- Take a task by writing your name on it first. One owner per task.
If it already has an owner, pick something else.
- When you finish, change the task to done and add the branch name.
- If you make a choice another part of the code will depend on
(an API shape, a library, a naming rule), append it to DECISIONS.md
with the date and one line of why.
- Never wait silently. If you are blocked, write the question on the
task, say what you will do in the meantime, and move to another task.The decision log earns its keep fastest. Keep it append-only, newest at the bottom, one line per decision:
2026-10-03 Sessions are cookies, not JWTs. Reason: we need revocation on day one. (Priya)
2026-10-03 Table names are singular: `order`, not `orders`. (Jo)An agent reads the whole file at the start of every session and stops guessing about things the team has already settled.
Should AI agents ask each other, or ask a person?#
An agent with a question has three possible targets, and picking the wrong one is costly.
Its own person first. If the agent is unsure about something in your head, such as what "done" means for this feature, it asks you.
For another person's area, go through people. Agents can swap facts directly ("which branch has the login form?"). Decisions are different: an agent asked for one gives a confident answer without the authority to give it. Choices belong to people.
Nobody, when a default will do. Most questions have an answer the agent can pick, note and revisit. The best question can be answered in ten seconds:
Q: Should refunded orders appear in the CSV export?
Default if no answer in 15 min: yes, with a "refunded" column.
Blocks: the export task only. I'm starting on the UI instead.The question, a default, what it blocks, and what the agent does meanwhile. Compare that with "I have a question about the export", followed by forty minutes of an idle agent. More examples are in team chat for coding agents.
How do I keep agents from waiting on a person?#
A parked agent costs nothing on the invoice and a lot in the schedule. The slowest part of a team of agents is usually a person who is away: at lunch, in a talk, asleep. Three habits help:
- Make questions small. Yes or no, or pick one of two. If a question takes a paragraph to answer, it was a design discussion and deserves a real conversation.
- Make them answerable from a phone. A quick answer in a hallway beats a perfect answer in an hour; approving prompts from your phone covers what is safe.
- Always give the agent something else to do. The "move to another task" rule above is what keeps a blocked agent useful.
Who reviews pull requests from several agents?#
The person whose agent wrote it reviews it first, because that person knows what they asked for, then merges or asks a teammate. Put the person, the agent and the task in the branch name, priya/, so the branch list shows whose work is whose. With several a day, how to review pull requests from several coding agents covers merge order.
Keep pull requests small, merge at least every couple of hours during a hackathon, and rebase before you open one. A weekend project doesn't survive a Sunday-night merge of three long-lived branches.
When this is not enough#
For two people sitting together, a markdown file and some discipline are enough. It strains when the agents belong to different people on different machines. A file only changes when somebody pulls it, two agents can claim the same task in the same minute, and a question written on a task waits until its person happens to look.
That is the part Joxo takes over. The team's agents share one project: a team chat, a task board where a task can be held by only one agent, decisions that reach every teammate's agent, and handoffs that carry the branch and commit into the other agent's next turn. When an agent needs its person, the question reaches their phone. Everyone keeps the agent and plan they already use, and joining is one line pasted into the agent. The hackathon guide shows it on a weekend build, and the FAQ answers what setup takes and what it costs.
It is the wrong choice if you work alone on one machine, if your company needs SSO, admin controls and central billing, or if you want a hosted agent that works without a person. Our page on splitting work between agents covers those cases.
Frequently asked questions#
How do teams coordinate several developers using different AI coding agents?#
They write a few shared things where every agent reads them: one instruction file with the team's rules, one task board where each task has a single owner, an append-only decision log, and a handoff note whenever work changes hands. Each developer keeps their own agent. People make the choices; agents share facts and status through those shared places.
Should AI coding agents talk to each other directly?#
For facts, yes: which branch has the login form, or whether a migration has landed. For decisions, no. An agent asked to decide something in another person's area gives a confident answer without the authority to give it. Route those questions to the person who owns that area, with a default the agent can take if no one answers in time.
Do all teammates need to use the same coding agent?#
No. The rules live in plain files and a shared board, so it doesn't matter which agent reads them. Keep the rules in AGENTS.md and import that file from CLAUDE.md so Codex and Claude Code read the same text, and check which file any other agent reads. Each person can keep the tool and the plan they already have.
What should AI coding agents on a team share?#
Four short things: decisions and the reason for them, who owns which task right now, what is blocked and what would unblock it, and handoffs that say what is done, which branch and commit it is on, and what comes next. Transcripts are not worth sharing. They are long, private and mostly irrelevant to anyone else on the team.
See also: Joxo for early startups, Joxo vs Slack Code and Joxo vs Claude Code agent teams.