# Don't Drop the BATON: Long-Horizon Robot Manipulation via Agentic Subtask Exploration and Transition-aware Memory

Source: [arXiv](https://arxiv.org/abs/2608.16889v1)  
Feed7 permalink: https://feed7.dev/p/2608-16889v1-17rof2d  
Published: 2026-08-17T17:59:57.000Z  
Trust: Needs Review (needs_review)

## Why Included

BATON explores robot subtasks independently, stores their solutions, and manages handoffs between them, replacing multiplicative whole-task search with additive composition.

## Source Summary

BATON freezes the vision-language-action model and explores each subtask separately, changing estimated exploration from **T^K to T×K**. Transition-aware memory checks invocation readiness, repairs handoff state, and selects outcomes compatible with the next stage without updating parameters.

## Practical Implication

The transferable agent pattern is to make subtasks independently testable and reusable, then represent their entry and exit contracts explicitly. That localizes failures and prevents a step that passes in isolation from leaving unusable state downstream.

## Agent-Ready Context

BATON freezes the vision-language-action model and explores each subtask separately, changing estimated exploration from **T^K to T×K**. Transition-aware memory checks invocation readiness, repairs handoff state, and selects outcomes compatible with the next stage without updating parameters.

The transferable agent pattern is to make subtasks independently testable and reusable, then represent their entry and exit contracts explicitly. That localizes failures and prevents a step that passes in isolation from leaving unusable state downstream.

On RoboMemArena, BATON reports **+11.6% task success** and **+14.9% cumulative success** over the prior state of the art. The evidence is specific to long-horizon robot manipulation, so applicability to software-agent workflows remains an analogy rather than a measured result.

## Connected Context

Feed7 judgment across 479 accumulated Signals:

BATON gives the prior harness guidance a more precise implementation pattern: independently exercise bounded subtasks, but treat each handoff as a checked state contract rather than assuming local success composes. Its robotics results support that architecture within one embodied domain; they do not establish the same gains for software or multi-agent workflows.

- [Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC Berkeley](https://feed7.dev/p/anthropic-s-cca-exam-as-a-field-guide-for-agentic-engineering-frank-coyl-00nf896) — BATON operationalizes the field guide’s bounded-subagent and output-isolation advice by checking whether each subtask’s resulting state is actually usable by the next.
- [Why We Killed Our Multi-Agent Pipeline — Subbiah Sethuraman and Abhilash Asokan, ZS Associates](https://feed7.dev/p/why-we-killed-our-multi-agent-pipeline-subbiah-sethuraman-and-abhilash-a-0fmz3z3) — The failed fixed pipeline identifies context loss and incoherent handoffs; BATON addresses the analogous composition failure with explicit transition readiness and outcome selection.
- [A Taxonomy of Cognitive Capability Gaps in Generative and Agentic AI](https://feed7.dev/p/2608-02553v1-12y8joy) — BATON supplies a concrete, empirically tested robotics mechanism for parts of the taxonomy’s durable-state, bounded-goal, monitoring, and adaptation requirements.

## Context Map

- Layer: agent
- Domains: None
- Topics: subagents, agent-memory, harness-engineering

## Uncertainty

- On RoboMemArena, BATON reports **+11.6% task success** and **+14.9% cumulative success** over the prior state of the art. The evidence is specific to long-horizon robot manipulation, so applicability to software-agent workflows remains an analogy rather than a measured result.

## 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.