Survey maps long-horizon agents into three layers, from in-context reasoning to lifelong experience

theomitsa · x · 2026-07-27

A survey frames long-horizon agents as a three-layer stack

The post shares a survey on long-horizon agents and an accompanying infographic. It argues that agent progress is being driven by a co-evolution between externalized harness engineering and internalized model optimization.

Key points

Externalized harness engineering includes workflows, memory management, tool interfaces, MCP, and verification hacks.
Internalized model optimization includes better core policies, synthetic fine-tuning, and agentic reinforcement learning.
The graphic breaks long-horizon agency into three levels:
H1: intra-context reasoning for tasks completed in a single window.
H2: cross-context memory for spanning hours or days with state and memory compression.
H3: cross-task experience for lifelong operation across open-ended streams, turning past interactions into reusable skills.
It also notes that the median task duration handled by frontier agents has grown from seconds in 2020 to many hours by 2026 in the chart.
Main failure modes are goal drift, context rot, and sparse rewards.
Example applications listed are software engineering, information seeking, and computer use.

Original post →

Survey maps long-horizon agents into three layers, from in-context reasoning to lifelong experience

More from coding & agent