Task and motion planning (TAMP) remains a hard problem even when the environment is fully observable and states are object-centric, because discrete decisions are tightly coupled with geometric, kinematic, and dynamic constraints. Generalized TAMP tries to ease this burden by exploiting regularities across problem instances, but existing approaches demand substantial TAMP-specific engineering.
A new research paper asks whether coding agents can automate that process by synthesizing programs that generalize across instances. Given a task description and access to a simulator, each agent decides how to interact with the environment while developing a program within a fixed synthesis budget. The program is then frozen and evaluated on unseen instances.
The authors evaluated Claude Code (Opus 5) and Codex (GPT-5.6 Sol and GPT-6 Astra) on 28 simulated environments from KinDER and PDDLStream, with object counts beyond those in the original benchmark. Across all program synthesis methods, they evaluated 980 generated programs on 100 held-out instances each, totaling 98,000 evaluation episodes.
All three agent configurations outperformed hand-engineered planners, one-shot generation, and an LLM-based generalized planning baseline in mean success: 56% to 95% versus 47% for the planners on the 16 environments where a planner is available. As object counts grow, the agents' programs maintained higher success than the planner while using an order of magnitude less computation per instance on average.
Logs indicate that agents used interaction to calibrate physical models, test edge cases, and refine strategies. The authors released all code, including the full prompts given to the agents, and conclude that coding agents are a strong baseline for generalized TAMP.