Kernel Architecture¶
Stable contracts¶
AgentModuleEngineStateSchemaDecision/ActionTask/Env/Memory
Canonical execution chain¶
DECIDE -> ACT -> REDUCE -> CHECK_STOP (loop)
prepare(state)is used inside DECIDE when model path is active.- no separate agent-level
observephase.
Architecture map¶
flowchart TB
T["Task"] --> R["Engine.run(...)"]
R --> D["DECIDE"]
D --> A["ACT"]
A --> X["REDUCE"]
X --> C["CHECK_STOP"]
C -->|"continue"| D
C -->|"stop"| E["END / FINALIZE"]
D -->|error| RD["RECOVER"]
A -->|error| RD
RD -->|"continue"| D
RD -->|"stop"| E
Responsibility boundaries¶
- AgentModule: policy semantics (
prepare/decide/reduce) - Engine: orchestration/runtime semantics
- ToolRegistry + Env: execution surface
- Hooks + Trace: observability side-channels
Minimal skeleton¶
See AgentModule API.