Hi everyone, I'm Alessio 👋 For the past months I've been building Certo, an open-source coding agent on the OpenHands Software Agent SDK. The SDK runs the agent loop and the tools; Certo adds the parts around it: which model runs, when to switch, and how to know the work is really done. What it does today: • Any model, one setting: Ollama, OpenAI, Anthropic, Gemini, DeepSeek, Kimi or anything LiteLLM supports • Picks a model per task (

--auto-model

), from ratings and task profiles you write • Escalation (

--escalate

): runs a task up a ladder of models, and hands it to the next one when a model can't finish or its provider fails, with a handoff note Certo writes itself • Independent verification: when the agent says it's done, Certo re-runs the project's own tests, build, lint and typecheck, with no LLM involved, and ends with one verdict and an exit code for scripts and CI • `certo bench`: compares models and strategies on tasks graded by hidden tests, and counts how often Certo itself was fooled The OpenHands SDK made a project like this possible for one person, so thank you to everyone working on it ❤️*.* I'd love to compare notes with anyone building model routing or verification on top of it, and I'm happy to share what I've hit along the way, especially with local models. https://github.com/alessioricco/certo

Amazing! We love people building on OpenHands, welcome to the community, and please share any experiences, pain points, etc!

Thank you! I'm looking for feedbacks as well, contributors and visibility are always welcome

Congrats @powerful-belgium-34200! Certo looks like an exciting project! I'd love to understand what stood out to you with the SDK that was very unique among other options you evaluated. Thanks again for sharing your story

Hi Joe, Thank you for your message OpenHands SDK was a big part of making Certo possible. Certo started as a proof of concept for an agent harness for an agentic AI stealth startup I’m advising. I was looking for a library that was powerful, model-agnostic, and complete enough to let me focus on building the agent rather than reinventing the wheel. I evaluated several options. With LangGraph, I would have needed to build more of the agent loop and supporting infrastructure myself. Between Pydantic AI and OpenHands, I chose OpenHands because it was specifically designed for building coding agents and already provided much of the infrastructure I was looking for. That let me focus on what really matters in Certo: selecting the best model for each task, defining when a task is actually complete rather than blindly trusting the model’s output, and setting the right execution limits. Certo has since evolved beyond that initial proof of concept into an independent open-source project, and I’m genuinely happy with the decisions I made along the way. Thanks again for building and maintaining the SDK!