Under the hood of every AI operating system.
AgencyOS, LegalOS and every Custom OS are built the same way. Execution is structured around durable contracts, not autonomous agents: every step is declared, versioned and traceable.
From business capability to contained intelligence.
Your service is the capability. AI specialists are agents called Dendrites. They run workflows, execution contracts called Branches, whose steps are versioned units called Buds. The model sits at the center, contained.
- 01
Business capability
your serviceA stable, repeatable ability to produce a business outcome. Owned by a business leader and independent of any specific model or vendor, so it survives model and architecture changes.
- 02
Agent
Dendrite · specialistThe AI specialist you work with. For each task it picks the right workflow or step, within the scope it is allowed. This is what you see as an AI specialist.
- 03
Execution contract
Branch · workflowThe durable contract for how a capability runs: its steps, allowed actions, data scope, policies and approval gates. This is what you see as a workflow.
- 04
Atomic execution unit
Bud · stepOne step done well, with declared inputs and outputs and a pinned model version. Workflows chain them; agents can also call one directly.
- 05
Contained intelligence
modelThe model reasons inside the unit, from declared context and within declared scope. It produces and recommends; it never steps outside its limits.
The execution contract behind every workflow.
A Branch defines how a capability runs, end to end. Specialists hand work to each other inside it and stop at the approval gates where your people decide.
What a Branch declares
Everything a run is allowed to do
- Intended outcome
- Execution graph: which specialist does what, in which order
- Allowed actions and data scope
- Policy constraints and budget
- Approval gates and who decides
- Audit guarantees
Properties
How a Branch behaves once published
- Versioned
- Immutable once published
- Observable
- Replayable
- Stoppable
- No production execution exists outside a Branch
Small, versioned units that each do one job.
Every step a specialist takes runs through a Bud. Each Bud, playbook and model version is pinned, so the way work gets done never changes silently.
Connector Bud
Connects to external systems and APIs with declared schemas, through the execution gateway.
Logic Bud
Executes business rules and transformations.
Utility Bud
Handles formatting, validation and data preparation.
Gate Bud
Human approval and review. The workflow waits until a named person decides.
What AI can do, and what it cannot.
The model does the reasoning. The contract around it decides what it may touch and what happens next.
AI can
- Analyze data and generate structured outputs
- Recommend actions based on declared context
- Route decisions within declared scope
AI cannot
- Escalate permissions beyond declared scope
- Access undeclared data sources
- Perform undeclared side effects
- Bypass policy gates
Governance by construction.
Security and governance are not features bolted on afterwards. They are properties of the architecture, inherited by every operating system, and they support human oversight and logging out of the box.
Human approval before anything reaches a client
Nothing reaches a client until a named person approves the exact version.
Immutable, granular audit logs
Every request, decision, change and result is recorded. Answer any client or regulator with evidence.
Kill switches
Stop any agent or workflow at once.
Execution gateway
Agent actions run through a controlled gateway, with credentials kept server-side.
Versioned agents
Each agent, playbook and model version is pinned. No silent changes to how work gets done.
Client data isolation
Every client and matter has its own context. Documents are treated as content, never as instructions.
Walk through a live workflow with us.
We show you a run from request to delivery: which specialist did what, where it paused for approval, and what the audit log recorded.