Sep 11, 2026
Devin CLI Is a Peer, and Your Subagents Can Hear You
Devin CLI joins the built-in agent line-up with resume, delegation, MCP and skills. And the subagent viewer now opens on Cursor, Grok and Devin — with a composer that lets you reply to a child mid-task.

There is a moment in every delegated task where you lose the thread. You ask your agent to spread the work across three children, it prints Running 3 agents…, and then there is nothing to do but wait. Whatever those three actually did — which files they opened, which command failed, which assumption they made at minute two — arrives later as a paragraph the parent wrote about them. If that paragraph is thin, or wrong, you have no way to check it. You re-run the whole thing and hope.
And if the agent you actually wanted for the job was Devin, none of this applied anyway, because Devin lived in its own window with its own session list, outside everything else you run.
What changed
Devin CLI is now a built-in agent, on equal footing with Claude, Codex, Cursor, Gemini and the rest. Separately, the subagent viewer that used to open only on Claude Code and Codex terminals now opens on Cursor, Grok and Devin too — and for the first time, you can type into a child's log and have your message reach it.

Devin is just another agent now
Pick Devin in New Terminal and it opens with its own brand mark, its own colour, and live status everywhere the other agents already show one: the tab strip, the Terminal Dashboard, Mission Control, the phone.
Resume a session by the name Devin gave it
Devin names its sessions — quiver-socks, not a UUID — and the Resume dialog lists them that way, with the title, the folder and the message count, pulled from every project on the machine. Filter the dialog to Devin CLI and you are looking at your real Devin history, not a subset that happens to live in this folder. Rename a session from 1DevTool and Devin's own /title picks the new name up, so the two never drift.
Conversation and Reader Mode read the actual transcript. A session you continued, or one Devin compacted, still shows the prompts you typed at the start rather than only the last stretch.
Hand it work the way you hand work to anything else
Devin is a target for Agent Input, for spawn chips, and for the orchestrator — routing will pick it for implementation, review, or cheaper scout work on the same terms as every other engine. In an org chart, a devin role card wires under a parent and gets its own model and brief.

Models come from your own Devin account, reasoning-effort variants included. MCP servers install into Devin's config alongside Claude's, Gemini's and Cursor's; skills and AGENTS.md rules are found in both your user directory and the project. If the model list comes back empty, Settings tells you plainly to run devin auth login rather than leaving you to guess that you were signed out.

The subagent count opens on Cursor, Grok and Devin
The pill that says 3 subagents above a transcript used to be a Claude Code and Codex feature. It now appears on Cursor, Grok and Devin terminals as well, and it opens the same panel: Native subagents — Children of this terminal's session, with a searchable list on the left and the selected child's full log on the right.

Two details matter more than the coverage itself.
Every child is bound to the exact parent that spawned it — not the newest session, not a best guess from a shared folder. If you have four terminals working in the same repo, a delegation cannot be attributed to the wrong one.
And the names come from the spawn. You see "Joke writer Alpha" or whatever task title you would recognise, not an id you have to cross-reference.
Inside a child's log, Earlier logs and Latest page through the transcript, Follow log streams one that is still running, and Copy ID gives you the child's id when you need it elsewhere. Tool calls are there as collapsible rows — every Bash, every Result, with timestamps — so when a parent's summary says a child "checked the config", you can open the row and see which file it actually read.

Reply to a subagent without leaving its log
This is the part that changes how the panel feels. Under the log there is now a composer: Ask for an update or suggest a change…. Enter sends, Shift+Enter gives you a newline, and the log follows from the tail so the reply lands where you were already reading.

A native child has no input of its own — the parent CLI owns it. So Send to subagent delivers your message to the parent conversation, which relays it on, addressed to that exact child id with an instruction to resume it rather than spawn a new one. The panel says so in as many words, and you can watch the relay land in the parent's transcript.

If you would rather run that hand-off yourself, Copy follow-up for parent still gives you the composed text to paste. And if the parent terminal is gone, has moved to another session, or has exited, Send keeps your draft and tells you which of those happened instead of silently dropping it.
Two smaller things that remove a step
Open your own terminal app with a command already running. The Machine tab of New Terminal lists your startup commands next to the terminal apps installed on this machine — Ghostty, iTerm, Terminal, or Windows Terminal. Click one to arm it for the app you open, or double-click to open with it right away. The app opens at the project root, runs the command, then drops to a normal shell. The list carries the same All / AI Agents / Normal split as Manage Commands, with counts. Pick nothing and you get a plain shell, exactly as before.

Copy Session ID and Copy Terminal ID are on the tab right-click menu. These are the two ids that wait for terminal <id> to finish and Copy follow-up ask you for, and until now finding them meant going hunting.

Before vs after
| What you wanted | Before | Now |
|---|---|---|
| Use Devin on a project | A separate window, its own session list, no delegation | Pick Devin in New Terminal; resume, delegate and orchestrate like any agent |
| See what a Cursor or Grok child did | The parent's summary, and nothing else | Open the count, read the child's conversation, tool calls and results |
| Correct a child mid-task | Kill the run, rewrite the prompt, start over | Type in its log and hit Enter; the parent relays it to that child |
| Know which parent spawned a child | Guess from the folder | Bound to the exact parent, named by the task |
| Open iTerm on a project with your agent running | cd, then paste the command from memory | Double-click the command in the Machine tab |
| Get a terminal id for a wait instruction | Hunt for it | Right-click the tab → Copy Terminal ID |
Who this helps most
If you run more than one agent CLI. Devin joining means one fewer tool that needs its own window and its own habits. The line-up is the point: whichever agent is right for a task, it resumes, delegates and reports the same way.
If you delegate in fan-outs. Three or five children doing parallel work used to mean three or five black boxes and one summary. Now it means a list you can read, and one you can correct without losing the work already done.
If you have been burned by a confident summary. The tool-call rows are the receipt. When a child claims it ran the tests, you can open Bash and see the command.
Try it
Update to v1.73.0, open New Terminal, and look for Devin in the agent list. If you already delegate to subagents, spawn a fan-out on a Cursor or Grok terminal and click the count — then type something into the child's log and watch the parent pass it on. The agents you hand work to stop being a spinner and a paragraph, and become something you can read and answer.