Pi Agent
A minimalist general-purpose coding agent with a custom-built harness focused on reliability, trust, and capability — built to run on a Raspberry Pi as a long-lived, locally-controlled executor.
What it is
Pi Agent is a coding agent built from scratch on a custom harness, designed to run as a long-lived executor on a Raspberry Pi. It is intentionally minimalist: the goal is not to chase frontier capabilities but to build a small, legible system where every tool call, every loop iteration, and every safety check is owned and inspectable.
Why it matters
Off-the-shelf agent frameworks optimize for capability and ergonomics — wrap a model in a tool-use loop, ship. That tradeoff hides what the agent is actually doing, and it makes trust opaque.
Pi Agent flips the priority order:
- Reliability — the harness is small enough to be fully read, debugged, and reasoned about.
- Trust — every tool call is bounded, observable, and reviewable. No hidden capabilities, no surprise side effects.
- Capability — added incrementally, only after the trust surface is solid.
Running on a Pi makes the constraints real. It can't lean on huge context windows or unlimited tool fanout — the design has to be deliberate about what gets loaded, what gets cached, and what gets handed off.
Status
Active development on a private repository. A future iteration will plug into GDDP as an executor adapter, giving the project graph a locally-controlled execution path alongside cloud agents like Jules.