Skip to main content

Architecture

Agent Vigilo is a distributed AI evaluation runtime built around durable state, explicit worker and coordinator ownership, isolated Wasm evaluators, and retry-safe event publication.

The runs direct commands and the and service modes. The owns global authority, while own shard-local work. RabbitMQ is the between coordinators and workers.

Core Views

System Structure

C4-style views of runtime boundaries, containers, components, and dependency relationships.

Performance Testing

Canonical noise calibration, bounded capacity, reviewed budgets, and confirmed regression gates.

Runtime Model

End-to-end execution model for agent calls, Wasm evaluator invocation, retries, aggregation, and event publication.

State Management

Database-backed lifecycle diagrams for runs, chunks, attempts, executions, outbox event records, and ownership guards.

Command Flows

Step-by-step diagrams for evaluator, run, coordinator, and worker command execution.

Failure Modes

Operational notes for duplicate delivery, worker crashes, stale attempts, resource exhaustion, and outbox retries.

Scaling

Logical run shards, single-cluster startup posture, and shard migration steps for larger PostgreSQL placements.

Runtime Limits

Configuration reference for Wasmtime memory, fuel, timeout, log, and worker semaphore controls.

  1. System Structure
  2. Runtime Model
  3. State Management
  4. Command Flows
  5. Failure Modes
  6. Scaling
  7. Performance Testing
  8. Glossary