Self-Programming AI Agents#

A self-programming AI agent can create or revise executable workflows while it works. In OpenProgram, those workflows use the @agentic_function decorator: the agent edits a source file with normal tools, the runtime validates and loads the function, and later turns can call it from the same registry as existing tools.

This is narrower than unrestricted self-modification. OpenProgram does not let a model silently replace the runtime or bypass validation. The editable unit is a reviewable function with a declared interface, explicit tool access, recorded execution, and normal source control.

What the agent programs#

An agentic function combines deterministic control flow with model decisions:

from openprogram import agentic_function


@agentic_function
def review_then_revise(draft: str, runtime=None) -> str:
    """Review a draft, then revise it against the review."""
    review = runtime.exec(
        content=f"Identify concrete defects in this draft:\n\n{draft}",
        toolset="none",
    )
    return runtime.exec(
        content=f"Revise the draft using this review:\n\n{review}\n\n{draft}",
        toolset="none",
    )

The Python body fixes the required order. The model handles the two semantic steps. Each call remains visible in OpenProgram's execution context.

How an agent creates one#

Install and start OpenProgram:

curl -fsSL https://raw.githubusercontent.com/Fzkuji/OpenProgram/main/scripts/install.sh | bash
openprogram

Then ask the agent for a bounded workflow, for example:

Create an agentic function named review_then_revise. It must review a draft,
revise it against the review, expose only its final output, and include a smoke
test. Show me the diff before committing it.

The bundled agentic-programming skill defines the file layout, decorator contract, validation steps, and smoke tests. A watcher can load an approved function without restarting the worker.

Runtime controls that still apply#

• Tool access comes from the function's explicit runtime call and configured approval policy.
• Function calls, model calls, and their context relationships are recorded in the session DAG.
• Resource limits, cancellation, structured output validation, and provider accounting remain runtime responsibilities.
• Source changes remain ordinary files: they can be inspected, tested, reverted, and reviewed before publication.

When to use it#

Use self-programming when a repeated task needs a reusable, inspectable process but the semantic steps still require model judgment. Use a plain function when the entire operation is deterministic. Use an existing tool when the operation already has a stable implementation.

Read the Agentic Programming guide, the @agentic_function reference, and the design rationale for the full execution contract.