Execution (agent-muscle)

In an AI system, “thinking” is safe, but “acting” is dangerous. The agent-muscle organ is the dedicated execution actuator for the Autonomic ecosystem. It is responsible for translating the decisions made by the agent-brain and agent-spine into real-world state changes on the host machine.

Structured Contracts

Traditional AI agents often run raw bash commands and attempt to parse the unstructured stdout using brittle regular expressions. agent-muscle rejects this approach.

Instead, every execution request sent to agent-muscle must specify a JSON schema contract. When agent-muscle runs a command (e.g., cargo test --message-format=json), it captures the output and validates it against the requested schema. If the output does not match the expected structure, the execution is marked as a failure, even if the process exit code was 0.

This strict contract ensures that downstream organs (like agent-spine workflows or agent-eyes visual QA) receive clean, predictable data payloads rather than unstructured text.

LoRA Fine-Tuning

Beyond executing shell commands, agent-muscle serves as the training actuator. It can orchestrate local LoRA fine-tuning jobs on Apple Silicon (via MLX) or CUDA GPUs (via Candle).

Before a training run begins, agent-muscle validates the training dataset against size constraints and formatting rules, ensuring that malformed data does not waste hours of GPU compute time.

Remote Actuation

Because agent-muscle subscribes to execution requests via the central agent-nerves event bus, it doesn’t have to run on the same machine as the rest of the ecosystem. You can run agent-brain locally on your laptop, while agent-muscle runs on a high-powered cloud VM, securely receiving encrypted execution payloads over JetStream.