How ceschmiedel/alm works
ALM is a governed federation of domain-specialist agents. A user submits an intent through the CLI or REST API; the runtime structures it as a LIP task, routes and decomposes it, executes the dependency plan with specialists, then arbitrates and synthesizes an answer with provenance. The Context Graph, scoped retrieval and governance support execution throughout. README documents evaluation, distillation, MLOps and Agentic Bus bridging; sampled excerpts do not establish all their runtime wiring.
Main entry points: runtime.py Federation runtime · models Model serving · experts Expert agents · cmrag Scoped retrieval
Components 20
Interfaces
- CLI
alm/cli.py - REST API
alm/api/app.py - LIP protocol
alm/protocol
Routing and execution
- Federation runtime
alm/federation/runtime.py - Cognitive router
alm/router/router.py - DAG execution
alm/orchestration - Context passing
alm/orchestration/context.py - Arbitration synthesis
alm/arbitration
Expert federation
- Expert agents
alm/experts - Model serving
alm/models - Model backends
alm/models/backends
Context and governance
- Context Graph
alm/graph - Scoped retrieval
alm/cmrag - IBAC policies
alm/governance - Session storage
alm/persistence
Learning and operations
- Evaluation harness
alm/evaluation - Distillation pipeline
alm/distillation - Model lifecycle
alm/mlops - Bus bridge
alm/bridge
Other components
- User
Connections 23
- User → CLI: submits intent
- User → REST API: submits intent
- CLI → Federation runtime: invokes
- REST API → Federation runtime: invokes
- Federation runtime → LIP protocol: structures task
- Federation runtime → Cognitive router: routes task
- Cognitive router → Context Graph: matches capabilities
- Federation runtime → DAG execution: executes plan
- Federation runtime → Expert agents: dispatches subtasks
- Federation runtime → Arbitration synthesis: synthesizes answers
- Federation runtime → Session storage: persists session
- DAG execution → Context passing: passes dependencies
- Context passing → Context Graph: publishes answers
- Expert agents → Model serving: requests inference
- Model serving → Model backends: dispatches inference
- Expert agents → Scoped retrieval: retrieves context
- Scoped retrieval → IBAC policies: checks access
- Scoped retrieval → Context Graph: uses ontology
- Federation runtime → IBAC policies: governs actions
- Distillation pipeline → Model serving: uses teacher serving
- Evaluation harness → Federation runtime: evaluates federation
- Model lifecycle → Model serving: manages versions
- Bus bridge → Federation runtime: publishes federation