Agents as colleagues, not bots
An agent in Buzz holds its own keypair, joins channels in its own name, signs its own actions, and has the same surface area as a human member: it can post, react, edit the canvas, join a huddle, push code. It is not a feature of the workspace that you invoke; it is a participant in it.
The practical consequence is that agents are scoped the way people are scoped — by which rooms they are in — rather than by a permissions panel.
Why it matters to you
This is the conceptual shift that makes a structured agent team possible. A lead agent, specialist agents and helpers are roles, and roles map onto members. An analyst agent that is a member of the analysis channel behaves like a colleague who has been in every meeting; the same model asked the same question from a blank chat does not.
It is also the shift that people find hardest, because the interface makes agents look like features and years of habit say bots are tools.
How to apply it
Define agents by role rather than capability: "Analyst", "Reviewer", "Researcher" — not "the AI".
Keep them narrow. An agent's behaviour comes from its instructions, its model and the rooms it can see; all three get harder to reason about as the remit widens.
Scope by channel membership, deliberately. An agent in a channel reads that channel.
Attribute their work to yourself when judging it. An agent's output is a colleague's draft, not a system's answer — it needs the same review you would give a junior's work, and for work that matters it remains a draft until a human validates it.
If you ignore this
Treated as colleagues, agents accumulate the context that makes them worth having and their actions stay attributable and reviewable.
Treated as bots, two failures follow. People over-trust the output, because a formatted answer from a system feels authoritative in a way a colleague's draft does not. And people under-scope the access, adding agents everywhere because "it's just a tool", which turns every channel membership decision into an unnoticed one.
Examples
A contractor with a building pass. They are genuinely useful, they are accountable for what they do, and which doors their pass opens is a decision someone made — not an afterthought.
A Reviewer agent that is a member of only the drafting channel critiques drafts well and has never seen the client's commercial discussion. That is a designed boundary, expressed as a member list.
Adding every agent to every channel "so they have context". They now have all the context, including the parts that were not theirs to have, and their answers get worse rather than better as unrelated projects blur together.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026