SwarmOS

The hard part of putting agents into a real business process is not getting them to do the work. It is being able to say, in advance and with confidence, what they cannot do.

SwarmOS separates the engine from the domain completely. A workflow is a bundle of declarations plus one behaviour module; the kernel walks the declared event graph, resolves each role to a model tier through a swappable provider profile, and enforces the bundle’s governance at every dispatch. Adding a new domain is authoring work, not engineering work. The engine does not change.

Governance is the product

Two constraints are enforced by the kernel rather than requested in a prompt.

Write-ownership. Each node declares the artifacts it may write. Anything outside that declaration is rejected — both when the node executes directly and when filtering the output of a live model. A step that processes one customer’s records cannot reach another’s. A step that drafts a document cannot alter the audit log.

Cost ceilings. Each workflow carries a spending limit evaluated against a live usage ledger during execution, not reconciled afterwards. When the ceiling is reached, the run stops.

Those two properties are what make agentic automation approvable inside an organisation that has something to lose. Most frameworks let you build the graph. Very few enforce anything once it is running.

One kernel, any domain

The same unchanged engine runs research synthesis, software delivery, document processing and operational workflows. Bundle libraries compose from open specialist agent collections spanning engineering, marketing, finance, security, project management, spatial analysis and more — which makes time-to-first-workflow days rather than months.

Every run emits a structured event stream, so the work agents did is legible to an operator without reading logs.

Deployed as the automation layer inside systems we build, or as a standalone workflow engagement.