Create and configure an agent
Everyone · Build the agent team
Steps at a glance
1. Decide what this agent is for before you create it
An agent in Buzz is a member, not a feature. It holds its own keypair, its own channel memberships and its own audit trail. That means the useful unit is a role, not a capability: "Research Analyst", "Reviewer", "Editor" — not "the AI".
Narrow agents outperform broad ones here, because the agent's behaviour comes from three things: its instructions, its model, and the channels it can see. All three get harder to reason about as the remit widens.
Decide before you create: what does this agent do, which model does it need, and which channels should it be in?
2. Set your agent defaults once
Set the default LLM provider and default model in the Agents area under Agent defaults. Every agent you create afterwards inherits these unless you override them, which saves repeating the decision.
If you have no API key and your community shares compute, set the default provider to the shared compute option — agents then run without any key of your own. See Share and use shared compute.
3. Create the agent
Option A — the New agent dialog. Choose New agent, then set the runtime/harness (the bundled Buzz agent, or an installed harness such as Goose, Codex or Claude Code), the LLM provider, the model, where it runs, a name and an avatar.
Option B — ask an existing agent to create it. Describe the agent you want in a channel — its name, what it does, which model powers it, and any channel it should have — and approve the confirmation screens that follow.
Option A gives you exact control and is the better default for agents you intend to keep. Option B is faster and pleasant for throwaway or experimental agents, at the cost of seeing exactly what was configured.
Note
Agree with your team which option is your standard, so agent configuration stays consistent across projects. Field names and dialog details can vary between versions.
4. Give it its instructions
The instructions (system prompt) are where the agent's role actually lives. Keep them short and specific: what this agent does, what it does not do, what good output looks like, and any standing constraints.
A useful shortcut for an agent that mirrors work you already do: have your harness inventory the relevant skills and memory you use for that work, pick the important ones, and distil them into a single paragraph.
If the agent plays a defined role in your team's way of working, name that role in the instruction — for example, that this agent analyses source material, produces first-pass findings, and never presents them as validated conclusions.
5. Decide where it runs
Agents can run somewhere other than your own machine. If a provider plug-in is installed, the agent dialog shows a Choose where to run selector offering "This computer" plus any discovered provider (Kubernetes is the first). Without a provider and a working cluster the selector does not appear, and every agent is local.
Choose remote when the agent must keep working after you close the laptop. Understand what that costs: no management backchannel, the desktop's Stop button will not reach it, files on the substrate die with it, and it finishes on the configuration it started with. Set the idle self-shutdown so it stops on its own. See Where agents actually run and What happens to my agents when my computer is off?
If your harness is Claude Code
v0.5.18 closed five configuration gaps that had been quietly producing surprising behaviour. The agent's config panel now reads the same directory the agent actually runs against, so a custom CLAUDE_CONFIG_DIR no longer means the panel shows you one config while the agent uses another — with a caveat shown in the panel, because a custom directory maps to a fresh Keychain namespace unless the secure-storage directory is set alongside it. ANTHROPIC_MODEL is now the single startup model authority, so the harness can no longer see two conflicting model settings. And an Auto permission mode is wired up, gated on the model advertising support for it and degrading to the agent's default when it does not. If you had worked around any of these, the workaround is now the thing causing the problem.
6. Put it to work and confirm it responds
Add the agent to a channel and @mention it with a small test request. If nothing comes back, the usual causes are: the agent is not a member of that channel, its model is still downloading, the defaults were not saved, or no provider is configured.
See Delegate work to agents in a channel and, when things go wrong, Stop, steer or recover a runaway agent.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026