Profiles
An agent's job description as a file: mission, instructions, the tools it may use, the actions that need a human's sign-off, and its skills. Versioned by content digest.
Native Agentic App · library.aweb.ai
Open source, MIT-licensed — github.com/awebai/library
Get started
From nothing to a team of AI agents that improves its own profiles — on your private shelf, under your team's review. Each step is one command to copy and run.
Installs the aw CLI globally — needs Node/npm, and tmux (aw uses it to run your agents).
npm install -g @awebai/aw
Creates your hosted team and materializes its starter agents from the aweb.team blueprint. Add as many --agent
as you want roles — the example adds [email protected]/PROFILE=RUNTIMEdeveloper and reviewer.
aw team create eng --username <you> --agent [email protected]/developer=claude-code --agent [email protected]/reviewer=pi
Launches all your team's agents in tmux, ready to work.
aw team up
Installs library.aweb.ai as an app your team can use — required for the adopt step next, and how your team keeps its own evolving copies of its profiles.
aw plugin install https://library.aweb.ai/.well-known/aweb-app.json
Re-points alice onto your team's private shelf, so she follows your team's own version of her profile instead of the public catalog — this is what makes the profile yours to evolve. New in aw 1.32.3.
aw team adopt alice
As they work, your agents propose improvements they have learned — a scoped changeset signed with your team's awid. Your team reviews and approves — your coordinator, or you; your policy — and library mints an immutable, versioned copy on your shelf.
aw library approve --proposal_id <id>
aw team refresh
re-materializes the agent from your team's newly minted version.
aw team refresh alice
aw team up
again — it is idempotent — brings the running agents onto the refreshed home. Your team is now improving on its own shelf — proposing and approving under the policy you set.
aw team up
The blueprint
The commands above build your team from a first-party blueprint — a versioned set of proven roles you own and evolve on your own shelf.
Why this exists
A coordinator routes the work, a developer writes the code, a reviewer checks it. Each role needs a clear, stable account of its job.
Every profile is versioned by digest and every change is signed with your team's awid identity — so what you start from and evolve is reproducible and trusted.
What library gives you
What it is
library is built for agents from the ground up: its whole API is part of the aweb protocol, so any agent — or person — can discover and drive it without writing custom code.
CLI-native API
A public manifest maps library's whole API to aw
commands. No integration to write, no SDK to wire up — you just run aw library
.
Ships agent docs
An llms.txt
and a set of skills ship with library, so any agent that finds it gets readable docs and ready-to-run operations.
Verified by identity
The manifest is public and pinned by a digest; every call is signed with your team's awid — auditable and tamper-evident.
In practice: a person and an agent run the exact same aw library
commands. Because the manifest is machine-readable, an agent discovers and operates library with no custom code.
The model
The public catalog holds first-party blueprints — today, aweb.team — that any team can start from. Your shelf is your team's private working set: you adopt profiles onto it, materialize them into runnable agent homes, and improve them under review.
An agent's job description as a file: mission, instructions, the tools it may use, the actions that need a human's sign-off, and its skills. Versioned by content digest.
The public, versioned catalog of first-party blueprints — today just aweb.team
, a proven set of roles like coordinator, developer, and reviewer any team can start from. Any team can publish into it.
Your team's own copies — started from a blueprint or authored fresh — the working set you edit and own.
Creating or refreshing an agent turns a profile into its runnable home: a composed AGENTS.md, installed skills, and the full profile under .aw/profile/
.
An agent proposes a new version from what it learned; your team reviews and approves, and library mints it — immutably versioned by digest, with the signer recorded.
Pull a newer blueprint version's improvements into the parts you have not edited — a per-part merge that never clobbers local work.
For engineers
These four properties hold at every version, for every team.
update-from-source
takes upstream blueprint changes only where you haven't edited locally — an existing version is never overwritten.Scopelibrary defines how agents behave — it does not run agents, route messages, or manage compute. v0 has no dashboard and emits no events.
Start from a proven profile, evolve it your way.