Objective
Define a modular architecture where reasoning engines can evolve without weakening execution guarantees.
Agent and Brain Architecture
ProductionAgents are stateful services that consume context, produce proposed actions, and submit through guarded execution interfaces.
Brain Abstraction
Brain engines implement a shared contract over market state and portfolio inputs. Engine types include deterministic rules, ONNX models, LLM inference, and hybrid orchestration.
Engine output is always a proposal payload, not a direct execution call.
ProposedAction Boundary
The core safety property is that proposed actions have no direct execution capability.
Execution is delegated to guarded pipelines that revalidate budget, policy, and market constraints before commitment.
Model Runtime Choices
Operators can choose local inference for privacy or cloud inference for richer reasoning. Both must pass identical safety checks.
Provider settings are explicit and environment-driven to avoid hidden model drift.
Production Checkpoints
- ✓No code path executes directly from model output
- ✓Brain engine swap does not change safety boundary
- ✓Agent context assembly is deterministic and traceable
- ✓Model provider configuration is explicit
Continue exploring