PhyAgentOS
A recursive self-improving agent operating system for embodied and robotic tasks, where an agent plans tool calls, a Gateway executes them, an observation collector captures before and after evidence, and a verifier checks the goal was actually achieved.
PhyAgentOS is an agent framework for embodied tasks built around a verified execution loop: the agent plans high-level tool calls, a Forge Tool API layer reports what its Gateway actually executed, an observation collector records before and after evidence of each action, and a task-level verifier judges whether the user-visible goal was really achieved rather than trusting the agent's own report.
Its execution architecture treats actions and their evidence as immutable and version-scoped records, so a skill's experience can be recovered and audited later, and includes a perception plugin system for configuring sensors with auditable environment writeback. Recent development has added independently installable and versioned Forge Skills bound into immutable task definitions, guarded skill evolution with scope-aware lessons, crash-safe SQLite-based orchestration, and support for the BEHAVIOR-1K embodied-AI benchmark.
It is written in Python and released under the MIT license.