rGi keeps an agent's work alive between prompts

What rGi is

An agent that only exists inside one prompt forgets the job the moment the prompt ends. rGi is a runtime that keeps the work alive: it records observations, asks a planner for actions, checks authority and budget before acting, executes what is approved, and retains the outcome.

When an outcome is ambiguous, execution stops for reconciliation rather than guessing. That is the whole design argument: a long running agent needs a brake, not just an engine.

What changed

This is a new project. Commit 8b29a16 implemented the perpetual runtime foundation with governed execution, and commit 2dd6adb added grounded adaptation, predictive planning and durable evaluation gates on the same day.

rGi describes itself as research on artificial general intelligence. Read it as a runtime experiment with governance built in, not a finished product.

Get started

A Rust toolchain and a checkout. The workspace test suite is the entry point:

cargo test --workspace --locked

Expected result: the workspace builds and the test suite passes.

There are Node and browser bindings if you want to drive the runtime from TypeScript:

cargo build -p rgi-napi --release --locked
cargo build -p rgi-wasm --release --target wasm32-unknown-unknown --locked

No documented MCP server. The supported routes are the Rust crates, the napi binding and the WASM build.

Use it today

Practical case: a long running maintenance agent. Input is a stream of observations from a system you operate. Workflow is observe, plan, authority and budget check, execute, retain. Output is a durable record of what was attempted and what it cost.

Acceptance test: give the runtime an action that exceeds its budget and confirm it refuses rather than executing and reporting the overrun afterwards.

Push it further

Experimental commentary. The ambitious version is an agent that runs for months and whose retained outcomes become the substrate for the next plan, with no human re-priming.

Limitation: nothing here proves long horizon stability, and the project is days old. Falsifiable test: restart the runtime mid task and confirm it resumes from retained state rather than starting over.