Waddle Labs

Waddle Labs (YC S26), which pitches itself as 'the Claude Code for robotics', builds LLM agents that control robots. A developer connects a robot to Waddle's API and types a task; the agents decompose the goal into subtasks and complete each one by watching camera feeds, writing control code and calling specialist models such as vision-language-action models, returning a program the developer can run and keep refining by talking to the agents. The company argues this combines the generalization of large end-to-end robot models with the reliability and explicit control of classical approaches, working across arms, grippers and camera setups without new data collection. Agents build and share a library of parametrized skills — a fold_grasp skill written to flip a package was later reused to fold a t-shirt — and a master agent can spawn subagents to coordinate several robots. Internally it reports creating working policies in about 20 minutes, having an agent collect around a thousand pick-and-place trials overnight and train a policy from scratch, and running overnight policy tuning with automatic scene resets; it says larger frontier LLMs with bigger thinking budgets consistently produce better policies. These are company-reported results. Founded in 2025 by CEO Hanming Ye and Yiding (Vincent) Song, who met at MIT's Research Science Institute and are roommates at Harvard; team size 2 per the YC directory. Access is via an early-access preview API. No funding amount or valuation has been disclosed beyond YC backing. Not to be confused with Waddle Labs, the US subsidiary of the Korean e-commerce AI company Waddle (Gentoo).