Xingyao Wang Portfolio
CS PhD candidate at UIUC / Co-founder of All Hands AI
CS PhD candidate at UIUC and co-founder of All Hands AI; lead creator of OpenHands (formerly OpenDevin), an open platform for coding agents that interact with a computer the way a human developer would, and of CodeAct, an ICML 2024 paper arguing agents should express actions as executable code rather than fixed-format tool calls.
Location
Didn't say. We're respecting the mystery.
Companies
Credentials
- UIUC
Skills
Notable
- CodeAct: ICML 2024 paper on agent action representation
- OpenHands: open platform for coding agents
What makes it work
A breakdown of the choices that make this portfolio stand out.
A specific, arguable position on how agents should act
CodeAct's central claim — that letting an agent write executable code as its action space outperforms constraining it to predefined JSON tool calls — is a concrete architectural bet, not a neutral survey of options, and it changed how several later agent frameworks structure tool use.
A benchmark for the failure mode that matters most in practice
MINT specifically evaluates multi-turn interaction with tools and language feedback, targeting the realistic case where an agent must recover from an initial mistake using feedback, not just succeed on the first try.
A research lab and a company built around the same architecture
Founding All Hands AI to build OpenHands on top of the same research (CodeAct, SWE-Gym) shows a direct, visible line from the academic contribution to a fielded product, rather than the two being disconnected.
What AI Agent Architects can take from this
Specific, actionable tips to apply to your own portfolio — no generic advice.
Take a specific, defensible position on a core architecture question (like how agents should represent actions) rather than presenting a neutral menu of approaches.
Design your evaluations around realistic failure recovery, not just first-try success — it's a harder and more useful bar.
When your research produces a genuinely better architecture, consider building the product around it yourself rather than only publishing the finding.
Projects
A research line that turned into a company. CodeAct argued agents should act by writing code, OpenHands built an open platform on that idea, and SWE-Gym trains agents on real repositories, all from a co-founder who is still publishing.
OpenHands (formerly OpenDevin)
An open platform for AI software-developer agents that work like a human developer: writing code, using a command line and browsing the web inside sandboxed environments. The current README presents Agent Canvas, a self-hosted control center that runs OpenHands, Claude Code, Codex or Gemini agents on local, remote or cloud backends.
He led the paper and co-founded the company behind it, taking an open research platform to a product with a very large open-source following.
Sandboxed code execution, terminal and web browsing for agents
Built-in evaluation across 15 benchmarks including SWE-bench
89.3k stars; 188+ contributors at paper time
CodeAct: Executable Code Actions Elicit Better LLM Agents
Proposes that agents act by writing executable Python instead of JSON or fixed-format tool calls. That gives them one action space in which tools compose and earlier actions can be revised after each observation. The team also released CodeActInstruct, 7k multi-turn trajectories, and fine-tuned CodeActAgent models.
Code as the action format is now a central design choice in agent architecture. He made the argument with a 17-model comparison, a dataset and trained models.
Python code as a unified, composable action space
Up to 20% higher success than JSON/text actions across 17 LLMs
Released CodeActInstruct (7k trajectories) and CodeActAgent models
SWE-Gym: Training Software Engineering Agents and Verifiers
An environment for training software-engineering agents: 2,438 real Python tasks, each with a codebase, an executable runtime, unit tests and a natural-language task. The team trained agents on it and also trained verifiers on sampled trajectories to scale inference-time compute.
Training agents, not just prompting them, is where coding agents are heading. This work provides the environment, the models and the verifier approach together.
2,438 executable, test-backed Python tasks
Agent fine-tuning plus trajectory-trained verifiers
32.0% on SWE-Bench Verified for an open-weight agent
More AI Agent Architect portfolios
Ready to build your portfolio?
Follow the AI Agent Architect roadmap — skills, projects, and timeline to get hired.