What is swarm?#
The one-line: an opinionated multi-agent ML platform for regulated industries.
The one-paragraph: swarm takes a problem statement in plain English, routes it through a team of specialized AI agents (problem-classifier, algorithm-selector, trainer, evaluator, deployer, auditor…), picks a model from an internal algorithm registry, trains it, runs compliance checks appropriate to your industry, and produces a signed audit PDF ready to hand to a regulator — in one afternoon instead of a two-week cross-functional ticket.
What swarm is not#
- Not a framework you build an MLOps stack out of. If you want that, look at Kubeflow or MLflow directly.
- Not a general-purpose agent orchestrator. LangGraph, CrewAI, AutoGen cover that. We borrow ideas; we're not trying to replace them.
- Not a chat assistant. The LLM is infrastructure, not the product.
What swarm is#
- A specialized, opinionated orchestrator for ML workflows with a compliance-first posture.
- A platform — REST + CLI + dashboard + plugin ecosystem.
- A certified-vertical-first play: RBI FREE-AI, HIPAA, EU AI Act conformity packs (each with regulator-format audit PDFs).
Where we fit in the landscape#
| Tool | Strength | Different from swarm |
|---|---|---|
| Kubeflow / MLflow | Pipeline orchestration at scale | No agent layer; DIY compliance story |
| LangChain / LangGraph | Generic LLM app framework | No ML training loop; no regulator-facing compliance bundle |
| CrewAI / AutoGen | Multi-agent orchestration | No ML training / deployment; no compliance |
| Databricks / SageMaker | End-to-end MLOps on cloud | Proprietary, expensive, not regulator-customizable |
| Anthropic Claude Code | Agentic coding assistant | Code-focused, not ML-focused; we're compatible with their plugins |
swarm sits at the intersection: specialized multi-agent orchestration + production ML training + regulatory evidence as first-class output.
Why regulated industries specifically?#
Three compounding reasons:
- High willingness to pay. A BFSI or healthcare org will pay $100K+/year for something that saves the compliance team a 2-week ticket per model.
- Regulation as moat. RBI directives, HIPAA clauses, EU AI Act Annex III are stable targets we can certify against. Competitors need to rebuild that certification path from scratch.
- Concrete buyers. Unlike "AI platform," a compliance-pack pitch has a named buyer (CRO / Head of Model Risk) with an explicit budget line.
Core design decisions#
These show up everywhere in the code; knowing them helps you navigate the rest of the docs.
- Composable monolith. One Python process runs the API + agent loop + scheduler. One Next.js process runs the dashboard. Event-log as the decomposition seam for later. Details.
- Permission engine is the mediator. Every tool dispatch, every HTTP route, every HITL gate resolves through one pipeline. Details.
- Plugins are CC-format compatible. A plugin that works in Claude Code works in swarm unchanged. Details.
- Compliance profiles are pluggable. RBI FREE-AI today; HIPAA and EU AI Act on the same extension surface. Details.
- Audit PDFs are tamper-evident. SHA-256 pinned, regulator-format layout. Details.
Status (April 2026)#
- Version: v0.11.0
- Tests: 628 passing, 0 failing
- License: Apache-2.0 (core) / commercial (compliance packs, managed hosting, SSO extras)
- Customers: design-partner program (pre-launch)
- Deployment modes: Docker Compose (dev), Kubernetes + Helm (v0.12), customer-VPC (v0.12), air-gapped (v0.13)
Next#
- Agents & teams — how the agent layer is organized
- Pipelines — how agents combine into workflows
- Compliance profiles — the buyer-facing differentiator