A 149-page survey says long-horizon agents depend on harnesses, not just bigger models
机器之心 · wechat · 2026-07-25
Long-horizon agents need more than long context
This 149-page survey argues that long-horizon agent capability is a system-level property, not just a stronger base model. It reviewed 900+ papers and engineering systems, and frames progress as the co-evolution of external harness engineering and internal model optimization.
What makes long tasks hard
- Small errors compound over long action chains.
- Context grows stale or gets corrupted as histories and tool outputs pile up.
- Sparse delayed rewards make credit assignment difficult.
- Irreversible actions raise the cost of mistakes.
The paper’s core framework
It separates the field into three task levels and three capability levels:
- H1 / C1: in-window interactive reasoning
- H2 / C2: cross-window or cross-session state and memory
- H3 / C3: cross-task-flow experience accumulation and generalization
Three phases of agent control
- Prompt engineering: steer the model through instructions and decomposition.
- Context engineering: actively manage retrieval, history, and summaries.
- Runtime harness: add workflow control, memory, tools, permissions, validation, recovery, and multi-agent orchestration.
Harness components
The survey breaks the harness into six pieces:
- workflows and planning/execution loops
- context and memory management
- tools, MCP, and skills
- orchestration
- hooks and middleware for policy enforcement
- verification and recovery
Internal optimization directions
It also surveys training-side work across:
- architectural support for long context
- trajectory and environment synthesis
- pretraining and intermediate training
- fine-tuning and curriculum learning
- agent RL
- on-policy distillation
- self-evolution
Why it matters
The paper’s main takeaway is that long-horizon agents will be judged less by how long they can run, and more by whether they can remain effective, economical, and trustworthy over long dependency chains. It highlights open problems in continuous learning, real-environment evaluation, budget-aware execution, and governance-by-harness.
More from AGI Musings
- Opus 5 reaches 30.2% on ARC-AGI 3 as critics question the benchmark — ChrSzegedy · 2026-07-25
- X debate says AI reviews could outclass many NeurIPS reviewers by 10x to 100x — peter_richtarik · 2026-07-25
- The next AI battle may be about memory lock-in, not benchmarks — VraserX · 2026-07-25
- Reuters asks whether chatbots will turbocharge cyberattacks and reshape security — wschroll · 2026-07-25
- AI will cheapen execution and make systems design the scarce skill — AryHHAry · 2026-07-25
- A short AI-community take says better models ship faster when guardrails take a back seat — victor_explore · 2026-07-25