01
Mission
One mission, one meta-agent. Each mission.md sets goals, boundaries and quality criteria, written in plain language and versioned like code.
mission.md
Architecture overview
How Constelio works
Fleets of agents that design, run and improve their own workflows: guided by a written mission, connected by an open message bus, running on your infrastructure.
The layers
From a written mission to work that gets done
02
Meta-agents
Each meta-agent reads its mission, then designs, tests and maintains the agents and workflows that fulfil it. This is the ADAS approach: Automated Design of Agentic Systems.
ADAS
03
Execution
Agents, workflows and plain deterministic code do the actual work. Language models are used where they add value; everything else stays predictable.
agents · workflows · code
04
Capability registry
Every agent publishes what it can do in a shared registry. Meta-agents look there first to reuse an existing capability before building a new one.
registry · discovery
05
Audit & traceability
Every call between agents, every model request and every decision is recorded with its context. You can trace any result back to the mission, the agent and the data behind it.
audit trail · call tracing
06
Fabric
One message bus carries messages, events and service discovery between all components. Git holds code, reviews and packages. Language models run on your premises.
NATS · Gitea · open-weight LLMs
Design principles
Built to be trusted in production
Separate fleets
No fleet maintains the infrastructure it depends on. Responsibilities are split so that a fault stays contained.
Everything is reviewable
Whatever a meta-agent builds or changes becomes a commit and a review. People approve what goes into production.
Fully traceable
Each action can be replayed and explained: who called whom, with what input, and why. Ready for internal control and audits.
People in the loop
The mission states which decisions stay human. Agents escalate instead of guessing.
Open and sovereign
A fully open-source stack and open-weight models, with no mandatory cloud. It runs on your servers.
The building blocks
- mission.mdOne per meta-agent
- NATSMessaging, events, discovery
- Capability registryWhat each agent can do
- Audit trailEvery call, traced
- GiteaCode, review, packages
- Open-weight LLMsOn-premise inference