Fu Zichuan

Fu Zichuan

AI: TRANSFORMING LIFE, REVOLUTIONIZING THE WORLD

OpenProgram — Open-Source Agent Harness

github.com/Fzkuji/OpenProgram · AGPL-3.0 · Python 3.11+ · macOS / Linux / Windows

The reference implementation of the Agentic Programming paradigm introduced in LLM-as-Code (KDD’26 AgenticSE Workshop).

One decorator turns a Python function into an agent: the docstring becomes the system prompt, type annotations become the tool schema, runtime.exec() calls become retryable DAG nodes, and plain if/for/return stays deterministic

Most agent frameworks make the LLM the orchestrator. OpenProgram inverts that: the Python program governs control flow, and the model is invoked only where a task genuinely calls for judgement.

@agentic_function
def triage(ticket: str, runtime=None) -> str:
    """Classify the ticket as bug / feature / question, then draft a reply."""
    kind = runtime.exec(ticket, choices=["bug", "feature", "question"])
    if kind == "bug":
        return runtime.exec(f"Reply using:\n{search_logs(ticket)}")
    return runtime.exec("Draft a short reply.")

One decorator binds what every other harness makes you hand-write — the docstring becomes the system prompt, type annotations become the tool schema, and plain if / for / return stays deterministic Python the model cannot skip.

Four mechanisms sit underneath:

  • Agentic Function — the primitive above; choices= is a code gate that sends an invalid answer back to be re-decided rather than letting it drift past.

  • DAG Context — every user turn, LLM call, and function call is one node on a flat DAG. Context is an addressable node rather than a per-agent buffer, so forking, spawning sub-agents, and cross-session messaging are all “select a different node set”.

    Every user, LLM, and function call is one node on a single flat DAG, tied by caller and predecessor edges; a branch can spawn a sub-agent on its own lane and attach the result back

  • Agentic Workflow — agents write and hot-load their own @agentic_functions with ordinary file edits; no create() / fix() machinery.
  • Event Infrastructure — one process-wide bus that the agent loop, auth, context, channels, and memory all emit onto.

Ships with a web UI, a terminal UI, and a library mode for embedding the engine in your own stack. Third-party harnesses install with one command.