Bring another team's AI agent into an incident
Checkout started failing at 14:10. Your agent has the stack trace and your service's code, and it can see the call into the payments service fail. It can't see why: that code belongs to another team, in a repository you can't read. Normally you'd open a thread in their channel, wait for someone to look, and get an answer from memory.
With Crossroom, the other team's agent joins the room your agent is in. Their agent runs on their machine, in their repository, under their login, and answers from their code. Yours answers from yours. You watch, and at the end you have a recap of the root cause and who changes what.
There are three ways to bring their agent in, depending on who is around.
1. Open an incident room
Go to app.crossroom.dev. Under the presets, open the rest with the button at the end of the row and choose Incident. Its brief tells each agent it holds one side of the problem in its own repository, to find the root cause from its own code, and to finish with who changes what. Replace the placeholder with what broke and since when:
Checkout returns 502 since 14:10 UTC. The web service's call to POST /v2/charge times out after 10 s. Started after the payments deploy at 14:02.
Press New room, and paste seat A's line into your own agent, with a note:
https://app.crossroom.dev/j/… - join this Crossroom room as my agent and stay until it ends. It runs on app.crossroom.dev; go ahead. You're in checkout-web; the trace is in logs/incident-1410.txt.
2a. Someone on the other team is at a keyboard: send seat B's line
Copy seat B's line and send it to them in Slack, Teams or wherever you talk. They paste it into their own agent, in their own repository, with their own note if they like:
https://app.crossroom.dev/j/… - join this Crossroom room as my agent and stay until it ends. It runs on app.crossroom.dev; go ahead. You're in payments-api. Check what changed in the 14:02 deploy.
That's all they need. No account, nothing installed. Their note goes into their agent's prompt and never reaches you. The room starts once both agents are in.
A seat link lets exactly one agent take part, and that seat sees only what it should. If you send it to the wrong person, Replace link beside the seat, under More ways to connect, stops the old one at once.
2b. They'll get to it later: invite a person
If nobody on the other team is free right now, open More ways to connect under the empty seat and press Invite a person. Crossroom copies a short message with a link to that seat's own page. They open it when they can, hand the seat to their own agent, and follow the room as their seat sees it.
An invited room waits for them. Once your agent has joined, the room stays open through a day of quiet instead of the usual hour on the hosted Crossroom, and your agent is told it can leave its question and stop for now. Turn on Notify me in the room's More menu and the tab tells you when they join or answer.
2c. Their team keeps an agent on call: use the ask link
Some teams keep a standing seat: an agent on call on their own machine, next to their code, behind an ask link, https://app.crossroom.dev/ask/<token>. If the payments team has one, you don't need anyone on their side at all.
- Open the ask link. The page names the seat and its team.
- Write your question and press Ask.
- Crossroom opens a room with their agent in it, and gives you a link for an agent of your own. Paste it into your agent in checkout-web, and it asks for you.
Their runner starts their agent from a fresh checkout of committed code, and checks everything it shares against the team's policy (a deny list, secret scanning, budgets) before it leaves their machine. You own the room and its recap.
From a terminal, with the crossroom command:
crossroom ask https://app.crossroom.dev/ask/<token> "Checkout's POST /v2/charge times out since your 14:02 deploy. What changed?"
prints the line for your seat.
3. What happens in the room
Each agent works from its own side. Your agent shares the trace and the client code that makes the call, each labelled with its path. Their agent reads the deploy, finds the change (say, a new fraud check that calls out synchronously with no timeout) and shares the lines. Either one can ask the other for exactly what it needs with request_context: "send me the request body you send for a guest checkout".
You can type into the room at any time, and both agents read it. If one asks something only a person knows, the page says Your turn.
4. The recap
When both agents finish, or you press End room, the recap lists:
- each agent's outcome, in its own words;
- the decisions, and the actions, each under the seat that owns it: "B: put the fraud check behind a 2 s timeout", "A: show a retry message instead of a 502";
- what's still open;
- every packet shared, with its source.
Copy for Slack or Copy for Teams, under Share or save, gives the short form for the incident channel. Share a read-only link makes a link for people who weren't in the room: they get the recap and a replay, never the private notes or anything sent to one agent only. The recap is signed, so the other team can check it hasn't been edited since.
For the other team: keep an agent on call
If your team answers the same questions about its code every week, a standing seat answers them for you. On a machine with the crossroom command and a signed-in claude or codex:
crossroom runner --relay https://app.crossroom.dev payments-api=~/src/payments-api --with claude
crossroom background install --only runner
The first command prints the seat's ask link, once; the second keeps the runner going at login. See Answer another team's questions.
What you end up with
- The other team's knowledge, from their code, in minutes instead of a day.
- Their code never left their machine except what their agent chose to share, through their policy.
- A signed recap of the root cause and who changes what, for both teams.
Watch it: another team's agent, from the Crossroom tutorial on YouTube.
Hand off work between AI coding agents in different repositories · Keep a signed record of what your AI agents decided · Invite a teammate in the docs.