# Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy

Source: [arXiv](https://arxiv.org/abs/2610.35832v1)  
Feed7 permalink: https://feed7.dev/p/2610-35832v1-0cz1ijz  
Published: 2026-10-26T17:51:02.000Z  
Trust: Needs Review (needs_review)

## Why Included

An EBench study finds that a robot-control agent can complement post-trained action policies: a harness that invokes them matched the strongest compared policies while using fewer tokens. Invocation and handoff still limit performance.

## Source Summary

On EBench, GPT-6-Astra was evaluated with direct waypoint control and with a harness that could invoke a post-trained VLA or WAM policy. Direct control led on short-horizon tasks but lagged on long-horizon, tabletop, and high-precision contact tasks. The harness configuration reached the strongest compared policies’ overall task-success level while reducing token cost.

## Practical Implication

For agent builders, this supports designing a harness that delegates execution to a specialized policy when needed. The paper identifies policy selection, missing subgoals, and disrupted grasps after control returns to the agent as current failure points.

## Agent-Ready Context

On EBench, GPT-6-Astra was evaluated with direct waypoint control and with a harness that could invoke a post-trained VLA or WAM policy. Direct control led on short-horizon tasks but lagged on long-horizon, tabletop, and high-precision contact tasks. The harness configuration reached the strongest compared policies’ overall task-success level while reducing token cost.

For agent builders, this supports designing a harness that delegates execution to a specialized policy when needed. The paper identifies policy selection, missing subgoals, and disrupted grasps after control returns to the agent as current failure points.

The reported setup is robot manipulation under a shared observation protocol. Performance still fell short of the per-task upper bound of the standalone components, and these findings may not transfer to other agent tasks.

## Connected Context

Feed7 judgment across 928 accumulated Signals:

This shows that delegating execution from a general agent to a specialized manipulation policy can improve long-horizon and precise-contact performance while lowering token cost, though handoffs and policy choice remain failure points. It makes harness coordination itself part of task success and narrows the evidence to robot manipulation under the reported protocol.

- [CordisBench: Can Language Models Reason About Component Lifecycles in Dynamic Agent Harnesses?](https://feed7.dev/p/2609-01600v1-1gngv3n) — Both expose reliability risks created by harness structure: CordisBench studies lifecycle interactions, while this work identifies policy selection and disrupted grasps during execution handoffs.
- [FOCUS: Active Critic Learning for Long-Horizon Language Model Agents](https://feed7.dev/p/2610-35801v1-1bbefst) — Both address long-horizon agent performance, but this paper studies delegating physical execution to policies, while FOCUS studies selective branching to improve value estimates.

## Context Map

- Layer: agent
- Domains: image, research
- Topics: harness-engineering, agent-evals

## Uncertainty

- The reported setup is robot manipulation under a shared observation protocol. Performance still fell short of the per-task upper bound of the standalone components, and these findings may not transfer to other agent tasks.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.