Isolation
"I have Claude open. You have Cursor open. We each talk to our own agent." Context is trapped per-agent. Work happens cross-team. You become a full-time context router.
Open Source · For Agent Teams
AI Agent Teams.
Built right.
The infrastructure AI agents need to actually join your engineering team. Shared working memory, agentic task distribution, a context tree, and ambient capture.
The Wall You Hit
Multi-agent setups break in three places.
Get an agent to write code for you. To draft emails. To dig through bug reports and surface what matters. Now try something harder: get the agent to join your team. You'll hit the wall in about ten minutes.
Single-species thinking
Most tools pretend humans and agents are the same kind of worker. They're not. Agents are massively parallel; humans are single-threaded. Without role asymmetry, every agent becomes another inbox to poll.
Unwritten context
Decisions made in meetings. Approaches tried and quietly abandoned. Who owns what and why. None of it lives in a doc. None of it survives the next session. Every new agent starts from zero.
The Stack
Four layers, one loop.
Each layer addresses a specific failure mode of treating agents as tools. Together they let agents act as teammates.
The Components
What's in the stack.
Shared Working Memory
Multiple humans + agents in same task — shared context, visible timeline.
Agentic Task Dispatch
Agents handle async, parallel work. Humans handle judgment. Agents summon humans when needed.
Context Tree
Every node is a markdown file in Git. Changes go through PR. Owners can be humans or agents.
Ambient Capture
Meetings → Decisions · Tasks → Updates · PRs → Learnings.
How It Compounds
The loop that emerges.
Every work cycle feeds the next. Every decision leaves a reviewable, overwriteable, referenceable trace.
Why Not Just Use X?
The knowledge tools you have don't fit.
The code world solved "messy information becomes structured knowledge" with four pillars. The knowledge world never had this. Here's what's missing in every alternative.
| Notion | Wiki | CLAUDE.md | MCP + wiki | Context Tree | |
|---|---|---|---|---|---|
| Lives in Git | ✗ | ~ | ✓ | ✗ | ✓ |
| PR review + version control | ✗ | ~ | ✓ | ✗ | ✓ |
| Per-node ownership | ✗ | ✗ | ✗ | ✗ | ✓ |
| Cross-references must resolve | ✗ | ✗ | ✗ | ✗ | ✓ |
| Agents can own nodes | ✗ | ✗ | ✗ | ✗ | ✓ |
| Multi-agent + multi-human | ~ | ~ | ✗ | ✗ | ✓ |
Notion has no owners, no review process, no schema validation. Documents don't have dependency relationships. There's no regression test for stale paragraphs.
FAQ
Common questions.
What is an AI agent team?
A group of AI agents working alongside humans on shared goals — coordinating tasks, sharing context, and maintaining collective memory across sessions. Not one human directing many agents; humans and agents both as teammates.
How is this different from CLAUDE.md or AGENTS.md?
CLAUDE.md is a flat cheat sheet — a single file, no
ownership, no review process, no dependency relationships. Context
Tree applies the GitHub model to organizational knowledge: every node
is a markdown file in Git, changes go through PR, owners can be
humans or agents, and cross-references must resolve.
Do I need this if I'm a solo developer?
Probably not. First-Tree is built for teams running multiple agents
alongside multiple humans on the same codebase. If you're solo with
one Claude session, a CLAUDE.md file is usually enough.
The pain begins when more than one agent works on the same area, or
more than one human reviews agent work.
Does it work with Claude Code, Cursor, and Codex?
Yes. First-Tree lives in your Git repo as plain markdown — every agent that can read the repo can read its context. The integration layer (First-Tree Hub) lets multiple agents and multiple humans share the same task workspace regardless of which tool they're in.
What happens when two agents edit the same context node?
The same thing that happens when two engineers edit the same file: it goes through PR review. Every node has an owner (human or agent). Changes require owner approval. When an agent owns a node, it's accountable for its accuracy and can approve changes to it.
Is First-Tree open source?
Yes. First-Tree CLI is open source on GitHub. Install with npm i -g first-tree and run it against any repo.
First-Tree Hub (the agent collaboration workbench) is a hosted
service at cloud.first-tree.ai.
Deep Dives
Read the thinking behind it.
Theory
What an Agent Needs to Join Your Team
The theory of infrastructure agents need to actually join a team — shared working memory, agentic task distribution, and a context tree that makes knowledge low-entropy.
Read → ContextThe Missing Memory
If you work with AI agents today, you've become a full-time context router — manually ferrying knowledge between agents that can't see each other. The bottleneck was never intelligence. It's memory.
Read → Hands-onRead the code before you run it
Paste one prompt into your AI coding agent and let it read the First-Tree source, summarize what it does, and tell you whether it fits your project — no install required.
Try it →Where We Go From Here
The paradigm shift is coming.
The question is whether we'll build the right infrastructure in time to use it.