Towards the Harness of Embodied Agents

6Related Work

6.1Coding Agents

[1]: Anthropic (2026), Claude Code by Anthropic: An Agentic Coding System. [2]: OpenAI (2026), Codex: AI Coding Partner from OpenAI. [43]: Wang et al. (2025), OpenHands: An Open Platform for AI Software Developers as Generalist Agents. [7]: Weng (2026), Harness Engineering for Self-Improvement. [44]: Yao et al. (2023), ReAct: Synergizing Reasoning and Acting in Language Models. [45]: Schick et al. (2023), Toolformer: Language Models Can Teach Themselves to Use Tools. [46]: Shinn et al. (2023), Reflexion: Language Agents with Verbal Reinforcement Learning. [47]: Packer et al. (2023), MemGPT: Towards LLMs as Operating Systems. [48]: Wang et al. (2024), Voyager: An Open-Ended Embodied Agent with Large Language Models.

Modern coding agents succeed at managing an environment end to end at scale [1, 2, 43]. Their shared design has been systematized as a harness [7]: a reactive loop that re-decides from the latest observed state [44], a tool registry the model selects from at runtime [45], recovery that reads a failure trace and retries [46], managed context [47], and memory that persists across sessions [48]. We transplant these patterns into the physical world, adapting each component along the way (§3). Two properties of the software environment have no physical counterpart. A code repository is readable, and its tests are verifiable by construction. The physical world provides neither, so we build both (§4).

6.2Embodied Foundation Models

[49]: Zitkovich et al. (2023), Rt-2: Vision-language-action models transfer web knowledge to robotic control. [50]: Kim et al. (2024), Openvla: An open-source vision-language-action model. [31]: Black et al. (2024), π0: A Vision-Language-Action Flow Model for General Robot Control. [10]: Black et al. (2025), π0.5: A Vision-Language-Action Model with Open-World Generalization. [51]: Ye et al. (2026), World action models are zero-shot policies. [12]: Berman et al. (2026), Claude Plays Robotics.

Embodied foundation models supply atomic capability in the physical world along two complementary lines. Vision-Language-Action (VLA) models map observation and instruction end to end to actions, from the RT series [49] through open generalist policies [50] to flow-matching architectures [10, 31]. World Action Models (WAMs) grow out of world models, predictors of how the world evolves, and fold action emission into the prediction network. DreamZero, a recent instance, drives a robot in closed loop from a single image-to-video diffusion model [51]. Both lines are the tools our harness calls, not the layer it replaces. A recent evaluation across control interfaces finds that frontier models mostly fail when driving joints directly but act capably when supervising pretrained policies, with the interface mattering as much as the model itself [12]. Today's best policies succeed on a single step with probability p of roughly 0.8 to 0.9 [10, 50], so an n-step task succeeds with only pn (§2), which collapses even for modest n; recovery and evaluation at the harness level close this gap without waiting for single-step perfection. This is why our design separates atomic capability from orchestration explicitly.

6.3Orchestration for Embodied Agents

[41]: Ahn et al. (2022), Do As I Can, Not As I Say: Grounding Language in Robotic Affordances. [52]: Liang et al. (2023), Code as policies: Language model programs for embodied control. [53]: Huang et al. (2022), Inner monologue: Embodied reasoning through planning with language models. [23]: Rana et al. (2023), Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. [54]: Duan et al. (2024), Aha: A vision-language-model for detecting and reasoning over failures in robotic manipulation. [55]: Bjorck et al. (2025), Gr00t n1: An open foundation model for generalist humanoid robots. [56]: Figure AI (2025), Helix: A Vision-Language-Action Model for Generalist Humanoid Control. [57]: Shi et al. (2025), Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models. [58]: Gemini Robotics Team (2025), Gemini Robotics: Bringing AI into the Physical World. [59]: Li et al. (2026), RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks. [32]: Fu et al. (2026), CaP-X: A framework for benchmarking and improving coding agents for robot manipulation. [60]: Liu et al. (2026), Guava: An Effective and Universal Harness for Embodied Manipulation. [61]: Xiao et al. (2026), ENPIRE: Agentic Robot Policy Self-Improvement in the Real World. [62]: Lu et al. (2026), ASPIRE: Agentic /Skills Discovery for Robotics.

Language-model orchestration over robot skills begins open-loop. SayCan selects affordance-grounded skills step by step but nothing judges the skill just executed [41], and Code as Policies emits a one-shot program [52]. Inner Monologue comes closest to closing the loop, threading environment feedback back through the language model, but the feedback channel is hand-picked and there is no general mechanism to decide whether a step has actually succeeded [53]. Later work supplies individual pieces of the loop, a scene graph the planner reads [23] and a failure judge that rules on outcomes [54]. Dual-system VLAs take a different route and pair a slow vision-language reasoner with a fast low-level controller [55, 56, 57, 58]. Most recently, concurrent systems bring coding agents to robotics: RoboClaw automates data collection through self-resetting action pairs [59], CaP-X benchmarks coding agents on manipulation and improves them by scaling test-time interaction [32], Guava searches the harness design space for manipulation and distills the result into a compact model for deployment [60], ENPIRE has coding agents self-improve policies on real robots through automated reset, rollout, and verification [61], and ASPIRE discovers reusable skills by writing and repairing control code [62]. Our work carries the paradigm of coding agents over as a whole. The harness is itself the deployed system, closing the loop from instruction to completion in the physical world.