Coding agents outperform hand-built planners on simulated planning tasks

single source· 1 articles · confidence: medium · first seen 2026-09-23 20:00 UTC

What this means for you

Nothing to buy or deploy — this is a benchmark result with code and prompts released. If you maintain a task-and-motion planner, the comparison you now have to clear is a coding agent given simulator access: 56–95% versus 47% mean success on the 16 environments where both were run, though on simulated tasks and without an evaluation date.

A preprint reports that coding agents (models that write and run code) can synthesise programs that solve task and motion planning problems — deciding what to do in what order under geometric, kinematic and physical constraints. Given a task description and simulator access, each agent worked to a fixed budget; the program was then frozen and run on unseen instances. Across 28 simulated environments from KinDER and PDDLStream, Opus 5 and two Codex configurations (GPT-5.6 Sol, GPT-6 Astra) reached 56–95% mean success on the 16 environments with a hand-engineered planner, versus 47% for those planners. 980 programs, 98,000 episodes. No evaluation date is given.

Models in this story

Key facts

  • ·Three agent configurations were tested: Claude Code with Opus 5, plus Codex with GPT-5.6 Sol and GPT-6 Astra. source
  • ·Agents reached 56–95% mean success against 47% for hand-engineered planners, on the 16 of 28 environments where a planner is available. source
  • ·980 generated programs were evaluated on 100 held-out instances each — 98,000 episodes — across 28 simulated environments from KinDER and PDDLStream. source
  • ·As object counts grow, the agents' programs maintain higher success than the planner while using an order of magnitude less computation per instance on average. source
  • ·The paper states no evaluation date; all code and the full agent prompts are released. source

What the sources say

  • Hugging Face Daily Papers — Reports coding agents synthesising planners from simulator access, with 98,000 evaluation episodes and released prompts.

Sources

The original reporting. Follow these — they did the work.

← the wire