Agent 5.0 is an open-source runtime for building local-first autonomous agents with deterministic decision routing, auditable memory, and safety-controlled execution.
It is designed to make agent behavior inspectable, reproducible, and privacy-preserving without requiring a cloud-based control plane.
Agent 5.0 follows a single deterministic runtime loop:
task → CognitiveKernelV2 → decision routing → MemoryOS → SafetyOS → execution → audit log
Provides a unified decision path for all agent actions and prevents scattered or hidden logic flow.
Stores all memory changes as append-only records for auditability, rollback, and traceability.
Implements a triple-review system (planner / critic / causal judge) to evaluate higher-risk actions before execution.
Constrains behavior using task context, state, and alignment signals.
Transforms repeated patterns into reusable strategy units for structured memory evolution.
Records key runtime events to ensure execution traceability and debugging transparency.
Agent 5.0 is currently an integrated prototype.
- Core execution loop is implemented
- Memory and safety subsystems exist at prototype level
- Audit logging and runtime tracing are functional
- System is under active engineering stabilization and refactoring
The project is not production-ready and is being actively developed toward release quality.
- Local-first execution (no required cloud dependency)
- Deterministic and inspectable decision flow
- Append-only memory and auditability
- Explicit safety control layer
- Minimal and modular architecture
- Whitepaper:
WHITEPAPER_EN.md - Architecture notes:
docs/golden_chains/
MIT License