OpenProgram Docs
Guide#
This directory is the home of OpenProgram's own programming model — the concepts you will not find in a generic LLM-framework tutorial, collected in one place. If you are writing functions for OpenProgram (or deciding whether to), read here first; the generic project docs (install, API index, troubleshooting) live in the other tabs. For the paradigm's background, see the philosophy.
Learning path#
Read in order; each step builds on the previous one.
| # | Doc | What it teaches |
|---|---|---|
| 1 | philosophy.md |
Why "agentic programming" — the rationale behind the model |
| 2 | writing-functions/agentic-function.md |
@agentic_function: wrap a Python function whose body runs LLM calls via runtime.exec(); composition patterns |
| 3 | writing-functions/function-metadata.md |
Parameter descriptions, placeholders, hidden arguments, render_range — the source of truth for function metadata |
| 4 | writing-functions/pure-python.md |
When NOT to use the decorator: plain deterministic helpers |
| 5 | embedding-in-your-own-stack.md |
Using the model as a plain library inside your own app or framework — bring your own LLM call, point state at your own directory |
Choosing the next step#
OpenProgram gives you three ways to decide "what runs next" inside a function. They are not alternatives to learn once and forget — picking the right one per task is the core skill:
| Doc | Mechanism | Use when |
|---|---|---|
choosing-the-next-step/fixed-order-calls.md |
Python code calls sub-functions in a fixed order | The step order is known ahead of time (pipelines: draft → review → revise) |
choosing-the-next-step/tool-calling.md |
Provider-native tool use: the model picks a function each turn, loop until it answers in text | Open-ended work where the model decides how many and which calls to make |
choosing-the-next-step/next-step-decision.md |
decision.make(prompt, options) / runtime.exec(..., choices=...): a text menu of options, the choice itself resolves into the next result |
Routing / finite branches; options may be plain values, not just functions; no provider tool-use support needed |
Reference#
../../reference/api/agentic-function.md— decorator API quick reference../../reference/api/runtime.md—Runtime.exec()parameters and behaviourskills/agentic-programming/SKILL.md— the in-repo skill an agent follows to write/validate functions (file layout, rule checklist, smoke tests)../../reference/design/function/calling-unification.md— internal design notes on the function-calling framework (evolution / refactor plans, not needed for authoring)- LLM-as-Code: Agentic Programming for Agent Harness (arXiv:2606.15874) — the paper describing the paradigm, accepted at the KDD 2026 AgenticSE workshop
Last updated · 2026-08-13