Coding agents outperform hand-engineered planners in simulated TAMP tasks
Coding agents using Claude Code and Codex synthesize programs for Task and Motion Planning that beat hand-engineered planners in mean success across 16 simulated environments.
On 24 Sep 2026, researchers released findings showing that coding agents can automate the synthesis of programs for Generalized Task and Motion Planning (TAMP). The study evaluated Claude Code (Opus 5) and Codex (GPT-5.6 Sol and GPT-6 Astra) across 28 simulated environments from KinDER and PDDLStream. The results are blunt: these agents achieved mean success rates between 56% and 95%, while hand-engineered planners only hit 47% on the 16 environments where a planner was available.
I work in AI and I build agentic AI systems, and my read is that the "engineering" in TAMP has been a bottleneck for too long. We usually see a trade-off between the flexibility of an LLM and the reliability of a formal planner. This data suggests that using an agent to write the formal program—rather than just predicting the next action—is the correct way to bridge that gap. It turns the LLM into a compiler for the physics of the simulator.
The agents didn't just guess. Logs show them using environment interaction to calibrate physical models and test edge cases. This is a critical distinction. They aren't just predicting a sequence of tokens; they are using a feedback loop to refine a strategy before freezing the program for evaluation on unseen instances. The researchers ran 98,000 evaluation episodes to verify this.
I don't know how these programs would hold up in a non-simulated, high-entropy physical environment. Simulation is a controlled variable, and the gap between a PDDLStream environment and a messy warehouse is wide. If the agents are relying on simulator-specific regularities, the "generalization" might be narrower than the success rates suggest.
Separately, the GRASP framework addresses the tendency of LLMs to collapse as task complexity increases. By decoupling the pipeline into GenPlan, RevPlan, and VerPlan, GRASP claims an absolute accuracy gain of up to 16.7% over direct LLM planners in interleaved dual-task environments. It also outperformed GPT-5-mini by 14.5%.
I think the "context-isolated modules" in GRASP are the real win here. Most people try to solve complex planning by stuffing more instructions into a single prompt, which just creates noise. Isolating the macro-guidelines from the local strategy exploration prevents the model from getting lost in the weeds of a complex trajectory.
Both these papers point toward a shift away from "prompting for a plan" and toward "architecting a process." Whether it is synthesizing a frozen program for TAMP or using a multi-stage discriminator in GRASP, the goal is to move the reasoning out of the latent space and into a verifiable structure.
