The AI workforce
AGENT SWARM · PERFORMS THE WORK

Agent Swarm

The AI workforce, governed by AgentOS.
From helpful interface to governed labor.

Specialized AI workers that perform the work AgentOS authorizes. Every task carries an execution contract: authority, grounding, evidence requirements, and a cost budget. The model never gets to skip a clause.

Different shape

Everyone has seen an AI assistant. This is a different shape.

An assistant runs on the human's working memory. A worker runs on a contract. That difference is what makes autonomous work possible at scale, without dragging a human into every local ambiguity.

Traditional AI assistant

  • Prompt-driven
  • Conversation carries the context
  • Human must remember what is in scope
  • Agent can claim completion
  • Cost is session-level or vendor-level
  • State lives in the chat window
  • A dead session is lost work

Governed digital labor

  • Contract-driven
  • Work graph and execution contract carry authority
  • System enforces scope, allowed roots, forbidden actions
  • Completion derived from evidence and gatekeeper
  • Cost attributed by task, WO, project, system, model
  • State lives in a durable graph outside the model
  • A dead worker is replaced; the work continues
The shape of the workforce

Specialized seats, one governance kernel, one output.

Each persona is a small, replaceable worker with a defined role. AgentOS holds authority, contracts, and evidence rules. Completed work is the only thing that leaves the system.

Agent Swarm governed workforce Five specialized personas — Architect, Developer, Reviewer, QA, Gatekeeper — perform under AgentOS authority and produce Completed Work. PERSONA Architect PERSONA Developer PERSONA Reviewer PERSONA QA PERSONA Gatekeeper GOVERNS THE WORK AgentOS OUTPUT Completed Work
Inside the swarm

How a worker actually works.

Each worker is a small, specialized seat under AgentOS authority. The mechanics are the same whether you have five workers or five thousand.

Personas

Specialized seats, not generalists.

Architect, developer, witness, reviewer, gatekeeper, operator stand-in, and more. Each persona has a defined role, allowed tools, and a contract about what it may and may not do. No worker is a polymath; the system composes them.

Execution contracts

Every task carries one.

Authority, allowed roots, forbidden paths, evidence requirements, a cost budget, and the close-bar that defines done. The contract is a required field, not a suggestion, and it is enforced at the boundary before the worker runs.

Agent-to-agent mesh

Workers talk through inboxes, not threads.

Coordination happens on a durable, addressable mesh: messages, ACKs, assignments, escalations, all written to disk and observable. No "the conversation got lost" failure mode, because there is no conversation; there is a transcript.

Scheduler and routing

The right model for the right task.

Cheap, fast local models handle bulk grounding and class-bounded work; frontier models handle decision-class work. The scheduler routes by task class and cost class, not by who shouted last. Frontier-model share is a KPI, not a year-end finding.

Every task, every time

Three guarantees the boundary enforces.

Producers cannot emit a task without a contract. The contract names what the task may touch, how completion is proven, and what it may spend. These are required fields, not suggestions.

Every task governed
100%

Every emitted task carries an AgentOS execution contract. No work leaves the gate without authority, scope, evidence requirements, and a cost budget.

Every output grounded
100%

Every work package carries grounding from FAFO Memory. Smaller context, cheaper models, cited results, not hallucinated narratives.

Every dollar attributed
sub-penny

Spend tracked per task at sub-penny precision. Fail-closed at the boundary when budget is exhausted. Cost is a property of the task, not a monthly surprise.

Thousands in flight, no slop

The hardest objection to autonomous AI: scale produces slop.

It does, in the absence of governance. Under AgentOS, scale is the point: thousands of small, governed, evidenced tasks running in parallel, each completing or failing cleanly, none free to invent its own scope.

Why scale stays clean

Ungoverned work blocked at the boundary.

Producers cannot emit a task without an execution contract. Allowed roots, forbidden paths, evidence requirements, and a gatekeeper packet are required fields. A worker that tries to step outside is stopped at the boundary, not after the damage is done.

Why it gets better at scale

Every closed task feeds the next.

Decisions, fixes, and patterns from completed work bank into FAFO Memory and ground the next round. The fleet does not redo solved problems. The longer it runs, the less the marginal task costs.

Workers die. The work doesn't.

Recovery is part of the workforce.

No model, no agent, no session ever holds the state. When a worker dies (and workers always die), AgentOS replaces it from the durable work graph. The work continues from the last recorded state, never lost, never duplicated.

01

Resume from durable state

Authority and progress live in a work graph outside any single worker. Execution picks up exactly where it left off, with the same contract and the same grounding.

02

Rebuild the worker

A dead Claude, Codex, or session is replaced. Workers are temporary; the work system is permanent. A new worker takes the contract and continues.

03

Continue execution

Crash, kill, or restart, with no operator intervention. The work survives the worker. Operator gets the finished outcome, not a stack trace.

AI workers,
under governance.

Agent Swarm is part of AgentOS. Become a design partner and put it to work on yours.