Thea: Towards the Harness of Embodied Agents

TL;DR

Thea (Eastern Institute of Technology, Ningbo) ports the coding-agent harness paradigm to the physical world: an agentic loop orchestrates robot capabilities, each wrapped as a callable tool, with core components inherited from coding-agent scaffolds (context engineering, skills, memory, safety hooks). The two abilities software grants for free but the physical world withholds — reading world state, and judging action outcomes — are bridged by two named primitives: Scene Graph as Context (a persistent symbolic world representation) and Evaluation as Exit Codes (per-action termination detection, success judgment, and on-failure cause diagnosis). Long-horizon behavior emerges from composing tools inside the closed loop between agent and world. Technical report only — no released code, benchmark numbers, or head-to-head model comparison at filing time.

Key claims

The harness is a design paradigm distinct from the model — for coding agents, achievement depends on the surrounding infrastructure (tools, context, memory, safety hooks) as much as the base LLM, and this generalizes to embodied agents [§Abstract].
Two abilities are trivial for coding agents but withheld from embodied agents: (a) reading world state — source code is text, but the physical world is only visible as a 30 fps RGB-D slice through time; (b) judging action outcomes — processes have exit codes and tests print PASS/FAIL, but a gripper closing on a cup vs. on air reports nothing back [§Thread posts 2–3].
Scene Graph as Context is Thea’s answer to gap (a): a persistent, symbolic representation of the world that accumulates across the agent loop rather than being reconstructed from raw frames per step [§Abstract].
Evaluation as Exit Codes is Thea’s answer to gap (b): a per-action module that (i) detects when an action should terminate, (ii) judges whether it succeeded, and (iii) on failure, diagnoses the cause — closing the software-style exit-code loop for physical actions [§Abstract].
Beyond those two primitives, Thea inherits from coding-agent scaffolds directly: context engineering assigns every input a lifetime (resident / refreshed / accumulated); skills load knowledge on demand; memory turns each task into per-tool experience; safety checks live in hooks [§Thread post 5].
Rich behaviors emerge from composition of tools within the closed loop, and this composition is what carries long-horizon tasks to completion in real environments [§Abstract].

Method

Thea is a harness: a fixed agentic loop that consumes an LLM and a set of robot capabilities-as-tools, and produces long-horizon behavior. The loop structure is inherited from coding agents. What is new is the physical-world adapter layer.

Scene Graph as Context replaces “read the codebase” with a persistent symbolic graph that the agent queries as if it were source code. The graph accumulates across steps — a slice of raw RGB-D isn’t the world; the graph is. Context engineering assigns each input a lifetime tier: resident (stays for the whole task), refreshed (recomputed each step), accumulated (grows over the trajectory).

Evaluation as Exit Codes is the physical analog of a process’s return value. For each tool invocation, Thea (i) detects termination (has the action’s target condition become true, or is it stuck?), (ii) judges outcome (did the grasp actually close on the cup?), and (iii) on failure, produces a structured diagnosis fed back into the agent loop as the “error message.”

Skills are loaded on demand rather than shipped as one giant prompt; memory promotes per-tool experience across tasks; safety checks are declared as hooks around tool invocations rather than baked into individual policies. The composition of these pieces — not any single one — is the harness contract.

Results

No headline benchmark numbers in the abstract or thread. The technical report is dated July 2026 and framed as a positional / architectural contribution.
Qualitative demonstrations only: the project page and Wentao Zhu’s launch thread describe real-environment long-horizon task completion driven by composed tool calls, with no comparison against baselines or ablation of individual primitives.
The concrete claim on the results side is emergence — rich behaviors from tool composition and closed-loop operation — rather than a metric [§Abstract].

Why it’s interesting

Thea joins a small but growing cluster the wiki is tracking under the coding-agent-as-robot-policy thesis: ENPIRE: Agentic Robot Policy Self-Improvement in the Real World (ENPIRE — frontier coding agents autonomously improve real-robot policies), Introducing Waddle: Agents that Control Robots (Waddle — deployed system where the LLM agent is the top-level policy and VLAs are callable tools), RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks (RoboClaw — VLM meta-controller with structured memory over VLA-as-tool), and RoboHarness: Memory-Driven Orchestration of Heterogeneous Robot Policies for Long-Horizon Planning (RoboHarness — memory-driven orchestration across VLAs, RL policies, and TAMP). What Thea contributes distinctly is naming the two gaps — reading world state, judging outcomes — that the coding-agent-to-robot port has to bridge, and proposing named primitives for each (Scene Graph as Context; Evaluation as Exit Codes). The Scene Graph as Context primitive complements SuperMap: A Spatio-Temporal SLAM System for Visual-Language Navigation‘s hierarchical 4D scene graph as text-queryable database, and Evaluation as Exit Codes is the “orchestrator deference” signal Anthropic tried to measure with a follow-rate metric in How Claude Performs on Robotics Tasks — the difference being that Thea builds it into the harness rather than measuring it post-hoc. The load-bearing weakness: this is a positional paper without benchmarks or a released harness at filing time, so it currently sits closer to Waddle’s demo-suite pole than to ENPIRE’s quantitative-eval pole in the same cluster.