Intel / article

Stirrup Open Source Framework Agents

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published Dec 10, 2025

Context

When Artificial Analysis measures how well a model handles agentic work, the score depends on the wrapped around the model: the loop, tools and handling it gets. This December 10, 2025 post introduces Stirrup, the open-source framework the firm built for that role and uses in its agentic benchmarks, including GDPval-AA, which it says was being released the same day (artificialanalysis.ai).

The design argument is that many frameworks are too rigid. According to Artificial Analysis, existing frameworks impose strict patterns that limit performance, while Stirrup steps aside and lets the model decide how to solve a multi-step task, in the way Claude Code drives its own workflow. The firm says it studied leading agent systems such as Claude Code for practical patterns in context handling, tool design and workflow stability and built them in. Developers still get structure: pre-built tools for web search and browsing, code execution locally, in Docker or in an e2b sandbox, an MCP client and document input and output; a generic interface for custom tools; automatic summarization to stay within context limits; and support for OpenAI-compatible APIs, including OpenRouter, and LiteLLM.

For readers of Artificial Analysis's agentic results, that makes Stirrup the environment behind the numbers: later write-ups, such as its Claude Opus 4.6 GDPval-AA result, describe GDPval-AA models working with shell access and web browsing in an through it. It is published as a Python package that installs with pip install stirrup, with documentation and a PyPI listing, and can be used as a package or as a starting template for custom agents.

Key takeaways · AI-distilled

Artificial Analysis pitches Stirrup on restraint: instead of imposing fixed patterns it says limit performance, the framework lets the model decide how to approach multi-step tasks, a design drawn from studying agents like Claude Code.
Bundled tools include web search and browsing, code execution, an MCP client and document input/output, and a generic Tool interface is meant to make custom tools simple to define and extend.
Model access runs through OpenAI-compatible APIs including OpenRouter, through LiteLLM, or a custom client; it handles image, video and audio inputs with automatic format handling and has built-in logging for debugging agents.

Terms in this piece · Glossary

agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agentic loop — The cycle an agent runs in: decide, call a tool, read the result, decide again — repeating until the goal is met or a stop condition fires.

Read the source artificialanalysis.ai