The reading list for agents with long-term commitment issues

What it does

This repository is the companion reading list to the survey Towards Long-Horizon Agents. It collects and categorizes papers exploring how large language models graduate from chatbots into autonomous systems capable of sustained reasoning, tool use, and revision across interdependent steps. The list mirrors the survey chapter-by-chapter, covering foundations, harness design, model optimization, and applications.

The interesting bit

The authors frame progress as a co-evolution between external harnesses—loops, memory, orchestration, verification—and internal policy improvements such as agentic reinforcement learning and self-evolution. They also formalize three nested difficulty levels (H1–H3) that separate intra-context reasoning from cross-session memory and lifelong task accumulation.

Key highlights

• Formalizes a long-horizon agent as Agent = π_θ ⊕ ℋ: a base policy coupled to a surrounding harness.
• Traces three evolutionary stages: prompt engineering (2020–2023), context engineering (2023–2025), and runtime harnesses (2025–present).
• Organizes research into two pillars: externalized harness engineering and internalized model optimization.
• Covers domains including software engineering, computer use, information seeking, and multimodal agents.
• Adopts METR’s empirical yardstick: measuring the task-completion horizon at a fixed success rate.

Caveats

• This is a curated bibliography, not a framework or codebase—its value is in curation, not compilation.
• The Chinese translation is machine-translated and only lightly reviewed; the English OpenReview version is authoritative.

Verdict

Worth bookmarking if you are researching agent architectures, training pipelines, or evaluation for sustained tasks. Skip it if you are looking for drop-in libraries or quick-start agent frameworks.