openai-agents-python

The OpenAI Agents SDK is a lightweight yet powerful framework for building multi-agent workflows. It is provider-agnostic, supporting the OpenAI Responses and Chat Completions APIs, as well as 100+ other LLMs.

Fork. This README is generated by the agentic harness and is fork-local — exclude it from any upstream pull request. Upstream docs are preserved in README.original.md.

Table of Contents

• Overview
• Quick Start
• Agentic Architecture
• Model Routing
• Project Layout
• Development
• Related Repositories

Overview

The OpenAI Agents SDK is a lightweight yet powerful framework for building multi-agent workflows. It is provider-agnostic, supporting the OpenAI Responses and Chat Completions APIs, as well as 100+ other LLMs.

Stack	python
Frameworks	fastapi, openai, pydantic
Tests	yes
Commits	1
Last activity	2025-10-07
Origin	fork of upstream

Quick Start

Install

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt   # or: pip install -e .

Run

# Entry point not auto-detected; inspect the layout below.

Agentic Architecture

This fork participates in the sahiixx agentic harness — a shared contract for how agents plan, act, verify, and recover across the account.

Fork note: this file is fork-local. Drop it from any branch you open as a pull request upstream. The harness contract also lives in AGENTS.md.

Signal strength: agentic density score 3288.

Patterns in play

Pattern	Role here
Prompt Chaining	Deterministic multi-step pipelines with known subtasks
Routing	Classify input, dispatch to the specialist path
Parallelization	Independent subtasks fan out; results aggregated
Orchestrator–Workers	Central planner decomposes dynamically
Evaluator–Optimizer	Generator/judge split with explicit rubric; bounded retry
ReAct	Interleaved reason → act → observe for adaptive tool use
Reflection	Self-critique before emitting a final answer

Reliability envelope

• Bounded execution — every loop has a max-iteration and wall-clock ceiling.
• Tool sandboxing — filesystem/network side effects isolated and reversible.
• Guardrail layering — validate at input, mid-loop, and output.
• Context engineering — select, compress, isolate; never grow unbounded.
• Self-verification — check intermediate output against constraints.

Model Routing

Agent work in this fork routes through Azure AI Foundry. See AGENTS.md.

Purpose	Deployment	Endpoint
Default / general	gpt-5.6-sol	/openai/v1/chat/completions
Deep reasoning	claude-opus-5	/openai/v1/responses only
Embeddings	text-embedding-3-small	/openai/v1/embeddings

export AZURE_FOUNDRY_API_KEY=...        # never commit this
export AZURE_FOUNDRY_BASE_URL=https://<resource>.openai.azure.com/openai/v1

Gotcha: Claude deployments on Azure return 404 api_not_supported on /chat/completions. They answer only via the Responses API.

Project Layout

AGENTS.md
CLAUDE.md
LICENSE
Makefile
README.md
docs/
examples/
mkdocs.yml
pyproject.toml
src/
tests/
uv.lock

Development

ruff check . && ruff format .
gh workflow run hermes-azure-check.yml

Secrets live in environment variables and CI secrets — never in tracked files.

Related Repositories

Part of a 226-repository workspace sharing one agentic contract:

• agentic-harness — patterns, contracts, reference implementations
• AGENTS.md in every repo pins identical model routing

---

_{README maintained by the agentic harness · fork-local · regenerated 2026-08-10}