Squads — how delegation works

When Anvil needs a specialist — a security reviewer, an infra expert, a compliance checker — it doesn't pick one by name. It names a capability, and the host chooses the best-suited specialist from its full roster. That capability is a squad.

Why it works this way

An AI model can only keep so much in its head at once. If every specialist agent — and there are dozens — had to be listed in the model's context, two things would go wrong: the list would crowd out the actual work, and it would grow every time you installed another agent, slowly starving the task.

Squads fix this. The model sees twelve squad names and a one-line description of each — a small, fixed cost of roughly 300 tokens that never grows no matter how large your agent library gets. The model asks for the capability it needs; the host quietly resolves that to a concrete specialist and runs it. The individual agents never enter the model's context at all.

The result: delegation that stays cheap and legible as your crew grows, and a model that reasons about what it needs done rather than who's on the bench.

The twelve built-in squads

Anvil began with the six technical squads; the six domain squads were added so the same delegation model serves work well beyond software.

How you use it

You rarely invoke a squad by hand — the orchestrator does it for you when a task calls for a specialist. When it delegates, it names the capability:

"Delegate this to the security squad: review the auth changes for injection and session-fixation risks."

Behind that, the host resolves "security" to the concrete agent best suited to the job and runs it in its own context, so the specialist's work never clutters the main conversation. You get the result back; you never had to know the roster.

Growing a squad from the marketplace

Here's the part that makes squads more than a fixed menu: you can add a thirteenth squad by installing a package — no new Anvil release required.

A marketplace agent bundle can declare its own squad (a name, a one-line brief, and its members). When you install it from the marketplace, that squad becomes routable immediately — the orchestrator can delegate to it exactly like a built-in. Install a package of medical-coding agents, and a medical-coding squad appears for the model to reach for.

Two guardrails keep this safe:

• Built-ins always win a name clash — an installed package can't quietly hijack security by reusing the name.
• Empty squads are never offered — a squad with no available specialist won't be presented as a choice.

Squads vs bots — not the same thing

A squad is a capability the orchestrator delegates to in the moment. A bot is a persistent member of your team — a named worker with a role, a model, and a ceiling it can't exceed. A bot serves a squad, but the two are different ideas: squads are about routing a task to the right skill; bots are about who's standing by to do the work. See Bots & the Team board.

Seeing it on AnvilHub

The Team page visualises the squad grouping of your crew — which bots serve which squads — so the delegation model you've been reading about becomes something you can look at.

Where to next

• Bots & the Team board — your persistent crew.
• Skills on demand — how Anvil finds and loads a capability only when it's needed.
• Marketplace — where new squads come from.