01
Canonical state
State is a deterministic projection of immutable events, not a loose memory blob.
Event-sourced runtime • persistent agents • governed autonomy
Runtime architecture for persistent AI agents.
Vigil is a governed runtime built for long-lived agent systems — event-sourced state, decision-first orchestration, replayable continuity, and hard runtime invariants.
state = reduce(events) → classify → decide → govern → act → replay
VIGIL SYSTEM MAP
Stable runtime
Event Truth
Replayable immutable history
State Projection
Deterministic reducer-based state
Decision Engine
Classify → decide → act
Governance
Risk and autonomy control
Continuity
Checkpointed recovery over event truth
Invariants
Hard runtime integrity guarantees
Runtime primitives
Vigil is not a prompt wrapper or agent shell. It is a runtime for persistent agent systems: event-backed state, decision-first control, invariants, and continuity over time.
02
Decision-first loop
Every turn routes through classify → decide → act, preserving control before generation.
03
Hard invariants
One-decision-per-turn, causation integrity, and state-event parity keep the system auditable.
04
Typed state domains
Relational, epistemic, strategic, operational, and task state remain explicit instead of collapsing into prompt text.
05
Governed autonomy
Autonomy is configurable and explicit, not an accidental side effect of prompts.
06
Continuity by design
Recovery comes from event truth and checkpoint discipline, not fragile session state.
Prompt pipelines optimize single turns.
Vigil is built for agents that must remain coherent over time.
Capability
Vigil
CrewAI
LangGraph
AutoGen
Event sourcing
✓
–
–
–
Crash replay
✓
–
–
–
Hard invariants
✓
–
–
–
Governance engine
✓
–
–
–
Autonomy levels
✓
–
–
–
Full audit trail
✓
–
–
–
Checkpoint & recovery
✓
–
◐
–
Multi-agent
◐
✓
✓
✓