What is DeepAgents in Glueprint?
DeepAgents is an open-source agent harness built by LangChain on top of LangGraph, distributed as the npm package deepagents and released under the MIT license. A single createDeepAgent() call returns a compiled LangGraph agent wired with four building blocks: a write_todos planning tool, filesystem tools (read_file, write_file, edit_file, ls, glob, grep), a task tool that spawns subagents with isolated context windows, and built-in prompt and middleware defaults. It is provider-agnostic, working with any tool-calling chat model.
Batteries-included AI coding and general purpose agent powered by LangChain.
DeepAgents connects over the open Agent Client Protocol. One generic adapter implements the protocol's client side, so DeepAgents gets the same surfaces as every other harness.
Everything below works with the DeepAgents CLI you have today — no migration, no new account for the model, no lock-in. Uninstall Glueprint and your terminal workflow is exactly as it was.
Set up DeepAgents on a host.
Installation is host-specific: repeat these steps anywhere this agent should execute. Portal and mobile can control the session afterward, but the process remains on that host.
Open the ACP catalog
On each execution host, open Settings -> Agents -> Browse ACP catalog and choose DeepAgents.
Install and initialize
Glueprint installs the registry distribution for DeepAgents, starts it, and completes the ACP initialize handshake.
Start a project session
Choose DeepAgents from a project or worktree and use the capabilities reported by that installed build.
How the connection works.
Glueprint adds orchestration and shared surfaces without taking ownership of the coding agent or its provider account.
Glueprint finds the DeepAgents binary you installed and spawns it, then speaks the Agent Client Protocol — JSON-RPC 2.0 over stdio — to drive it. Same CLI, same authentication; Glueprint just wraps the session.
Message chunks, tool calls, and tool-call updates from DeepAgents arrive as protocol notifications and render as the same live, typed timeline every Glueprint session gets — watchable from desktop, browser, or phone.
When DeepAgents asks to write a file or run a command, the protocol's permission request surfaces as an inline approval card on whatever device you're holding. Approve, deny, or auto-allow per tool.
Because DeepAgents speaks ACP, adding it is a catalog entry, not an engineering project — the same adapter runs every other agent in the registry, so you can tile them side by side.
Capabilities and discovery.
The status labels separate guaranteed integration behavior from features owned or announced by the installed agent.
| Capability | Status | Behavior |
|---|---|---|
| Built on LangGraph, MIT licensed | Runtime | Confirmed when the installed agent connects. |
| write_todos planning tool for task breakdown | Runtime | Confirmed when the installed agent connects. |
| task tool spawns isolated-context subagents | Runtime | Confirmed when the installed agent connects. |
| Filesystem tools as working memory | Runtime | Confirmed when the installed agent connects. |
| Provider-agnostic, works with tool-calling models | Runtime | Confirmed when the installed agent connects. |
Authentication, ownership, and limits.
Glueprint is the environment around the agent. Credentials, model access, and vendor-specific behavior stay where their owner designed them.
Watch every DeepAgents prompt, tool call, and diff stream in real time — desktop, web portal, iOS and Android. The session runs on your machine; the view follows you.
Permission prompts arrive as approval cards wherever you are. Kick off a refactor at your desk, approve the file writes from the couch.
Run several DeepAgents sessions side by side, each in its own git worktree, with diffs, history, and a three-pane conflict resolver alongside.
Tile DeepAgents next to Claude Code, Codex, and the rest of the ACP fleet in the Command Center. The protocol differs; the surface doesn't.
Promote it to an assistant: persona, constitution, long-term memory, and a daily journal on disk, loaded into every wake.
Wake it on cron: triage the inbox every 30 minutes, sweep PRs at 9am, run the nightly cleanup. Quiet hours and watermarks prevent re-fires.
Messages, sessions, and output tokens computed from session logs — compare DeepAgents against Claude Code, Codex, and your other ACP agents.
Remote viewing rides AES-256-GCM envelope encryption. The relay routes; it cannot read.
Common questions
Can I run DeepAgents with Glueprint?
Yes. DeepAgents is registered in the official Agent Client Protocol registry, and Glueprint implements the protocol's client side. It spawns the DeepAgents CLI you already installed and drives it over JSON-RPC 2.0 — adding live sessions, inline approvals, worktrees, and remote control.
Do I need a separate DeepAgents subscription or API key?
No. Glueprint doesn't resell model access. DeepAgents authenticates exactly the way its vendor designed — its own login, subscription, or API key. Glueprint spawns the binary and speaks the protocol; it doesn't sit between you and the model provider.
Can I run DeepAgents from my phone?
Sessions always run on your machines — a laptop, a desktop, or a server with the headless daemon. The iOS and Android app streams DeepAgents sessions live, lets you send follow-ups, and handles permission approvals from anywhere.
Can I run DeepAgents next to Claude Code and Codex?
Yes — that's the point. Each harness is an adapter; sessions from all of them tile into the same Command Center, share the same task boards, and roll up in the same per-harness analytics.
Is my session content private?
Yes. Remote viewing routes AES-256-GCM-encrypted envelopes through the relay, which stores a wrapped key it cannot open — decryption happens on your devices. On the free local tier, session content never leaves your machine at all. See security.