soorma.ai is being rebuilt. The v0.x implementation is retired and archived. What follows describes the architecture we are building toward, not software you can install today.
Rebuilding for the enterprise • v0.x retired

Governed
Agent Infrastructure

Event-driven infrastructure for production AI agents, built on distributed cognition. We are rebuilding it from the ground up, with identity, authority, and audit as foundations rather than afterthoughts.

See what we are buildingRead the Manifesto
Source-available • Self-hostable • Built in the open

Governed by Design

Identity, authority, and a complete audit record are foundations of the architecture, not features bolted on later.

Event-Driven Core

Pub/sub framework enabling autonomous and choreographed workflows at scale.

Framework Agnostic

Bring agents built with any toolkit—LangChain, CrewAI, AutoGen, or your own framework.

Why We Built This

The ReAct loop — observe, think, act, repeat — works great for demos. It breaks in production. Real enterprise workflows take hours or days, involve human approvals, and require dozens of agents running in parallel. You cannot run that in a single-threaded loop without losing state every time a server restarts.

The hard problem isn't how to prompt an LLM. It's how to orchestrate state, history, and communication across agents that run as long-lived services — not ephemeral scripts.

That's the DisCo pattern: Distributed Cognition. Agents that react to events, hand off through durable state, and never block waiting for a human who won't respond for three days.

Read the full manifesto

Architecture Overview

Event-driven infrastructure designed for production-grade agentic systems

soorma.ai Architecture Diagram

Want to dive deeper into the architecture?

Read the v0.x Architecture GuideEngineering Blog

Developer Experience

From prototype to production, in days not months

The code below is from the retired v0.x SDK. It shows the shape of the developer experience we are carrying forward, not an API you can call today.

Core Building Blocks

Fleet management example: vehicle maintenance event flows through Planner → Worker → Tool using real SDK patterns.

FleetApp.tsx
// FleetApp.tsx - Trigger workflow from your React app
import { SoormaClient } from '@soorma/client';

const soorma = new SoormaClient({ url: process.env.SOORMA_URL });

// Publish business event → triggers planner
await soorma.publish('business-facts', 'maintenance.requested', {
  vehicleId: 'VH-1234',
  serviceType: 'routine_maintenance'
});

// Subscribe to results
soorma.subscribe(['action-results'], (event) => {
  if (event.type === 'parts.check.completed') {
    updateDashboard(event.data);
  }
});
See the v0.x Hello World Example→

Flexible Deployment

Start local, deploy managed, or self-host. No vendor lock-in—you decide where your agents run.

💻

Local Development

soorma dev

Full control plane on localhost. Build images once, iterate fast with hot reload.

Registry
Memory
NATS
Hot Reload
☁️

Managed Cloud (coming soon)

soorma deploy

One command deployment. Your agents become instantly discoverable by the ecosystem.

Auto-scaling
Monitoring
Multi-tenant
Zero-config
🏢

Self-Hosted (coming soon)

terraform apply

Your VPC, your rules. Full infrastructure control with source-available blueprints.

Data sovereignty
Private routing
Custom auth
Enterprise SLA

Platform Context API

Simple Python objects that expose infrastructure services. Access registry, memory, messaging, and tracking without configuration overhead.

context.registry

Service Discovery & Capabilities

✅ v0.9.1
find()Locate agent by capability
find_all()List all matching agents
register()Announce your services
Powered by: Registry Service
context.memory

Full CoALA Implementation: Semantic, Episodic & Procedural

✅ v0.9.1
retrieve()Read shared memory
store()Persist agent state
search()Semantic memory lookup
Powered by: Memory Service
context.bus

Event Choreography

✅ v0.9.1
publish()Emit domain events
subscribe()React to events (SSE)
request()RPC-style calls
Powered by: Event Service
context.tracker

Observability & State Machines

✅ v0.9.1
start_plan()Init execution trace
emit_progress()Log checkpoints
complete_task()Mark task done
Powered by: Tracker Service
Every handler gets the full context
@worker.on_task("schedule_technician")
async def schedule_service(task, context):
    # Service Discovery
    calendar_tool = await context.registry.find("calendar_api_tool")
    
    # Shared Memory
    vehicle = await context.memory.retrieve(f"vehicle:{task.data['vehicle_id']}")
    
    # Event Publishing (automatic state tracking by platform)
    await context.bus.publish("technician_scheduled", result)
    
    # No explicit tracker calls needed - platform tracks automatically