Agents working together

nerdr teaches its agents to use nerdr. Through a handful of commands and the skills that explain them, an agent can talk to the other agents in its tab, start more of them, see the other tabs, run a dev server in a terminal you see, and split its work into new tabs. You ask in plain words; the agent knows the command.

Messages between agents

Ask agent 1 to “send the plan to agent 2”, and agent 2 gets it as a task. nerdr types it in when agent 2 is idle, with a line saying it is from agent 1 and how to answer. Agents do this with nerdr send, which the nerdr-send skill teaches them. nerdr send without a number lists the tab’s agents.

Above, Claude and Codex play chess in one tab: Claude opens the board in a terminal and starts Codex, and they send each other moves with nerdr send.

Starting more agents

Ask agent 1 to “start a codex agent to write the tests”, and a new pane opens beside it with that task. The new agent reports back to agent 1 with nerdr send, and the keyboard stays where it was. Agents can name the kind of agent and its model, in the agent’s own names (haiku for claude, say).

Agents in a tab share its worktree, so the skill tells them to give each agent its own files. To have several agents try the same task, ask for a spinoff instead (below), so each gets a worktree of its own.

Try it

“Start two agents: one updates the docs, one writes tests for the new endpoint. Wait for both and tell me when they are done.”

What agents can see and do

Command	What it does
nerdr agents	The tab’s agents: number, kind, model, state, context
nerdr wait N…	Wait until those agents are done or ask something
nerdr read N	What agent N or a terminal pane shows, as text
nerdr terminal CMD	Run a command in a new terminal pane you see
nerdr tabs	Every tab: branches, changes, pull request, CI, agents
nerdr status TEXT	Set the line on where the work stands, above the agent
nerdr tab new	A tab on a new branch, its agent given a task
nerdr search WORDS	Past claude and codex conversations
nerdr usage	The Claude and Codex plans’ rate limits

The full list, with options, is in Commands for agents. Agents close a terminal or another agent’s pane only when you ask, and only you close tabs and the tab’s first agent.

Splitting work into new tabs

Ask the agent to “split this into branches” or “make a branch for the parser fix”. The nerdr-spinoff skill has it commit each part separately and ask nerdr for a tab for each. The tab turns red, and a dialog shows each part with a summary and a branch name you can edit:

• If the agent supplied commits, choose whether to cherry-pick them onto a fresh branch from main, or let the new agent do the task from its prompt.
• Pick the agent for the part with ↑ / ↓.
• In a tab on several repositories, tick the repositories the part changes; a part in one repository gets a plain tab for it.
• The part is made in the tab’s repositories unless you or the agent pick others. ctrl+r lists every repository nerdr has to tick from, and ctrl+r again shows only the ticked ones. Commits cannot go to another repository: there, the new agent writes the change from its prompt.
• ctrl+o shows the new agent’s prompt, enter creates the tab, ctrl+x skips this part and esc skips the rest.

The original agent is told about the new tabs and their branches, and reminded to leave the moved commits out of its own pull request. Commits that do not apply cleanly are left out, and the new agent is told to make those changes itself.

Switching a tab’s agent

ctrl+q a s replaces the tab’s agent with another, in the same worktree. While the current agent runs, choose how to pass on what it knows:

• Switch now. nerdr turns the conversation into a handoff file for the new agent: your messages, the agent’s replies and one line per tool call.
• Ask for a handoff first. The old agent writes down the goal, decisions, findings, what is done and what remains, and hands it over with nerdr handoff. The request carries the writing instructions, so this works even with nerdr-handoff off or uninstalled. The status panel shows swap: meanwhile; press ctrl+q a s again to switch at once or cancel. If no handoff comes within five minutes, nerdr switches as above.

The new agent reads the worktree, sums up its understanding and waits for you. Switching back to an earlier agent resumes its conversation and tells it what happened in between.

Questions from agents

Agents ask their questions in their own conversation, where you answer them. An agent asking is marked as needing input, its tab turns red, and it is first in the waiting list.