A survey maps the real bottleneck in long-horizon agents
青稞AI · wechat · 2026-07-21
This long-form article summarizes a 149-page survey on long-horizon agents and argues that the key bottleneck for production agents is not just single-step intelligence, but sustained execution over long dependency chains.
Main idea
Long-horizon ability is defined by how well an agent can keep goals intact, manage context, recover from errors, and finish tasks over long stretches of interaction. The survey frames this as a joint evolution of:
- External harness engineering: workflow orchestration, memory, tools, skills, hooks, permissions, and verification.
- Internal model optimization: architectures, data synthesis, pretraining, finetuning, agent RL, distillation, and self-improvement.
Key framework
The article explains a three-level ladder:
- H1/C1: within-window interactive reasoning.
- H2/C2: cross-window and cross-session memory.
- H3/C3: cross-task experience accumulation and continual improvement.
It also traces the system shift from prompt engineering to context engineering and now to runtime harnesses that control tools, memory, workflows, validation, and recovery.
Why it matters
The piece argues that long-horizon agents fail mainly through goal drift, context corruption, sparse delayed rewards, and irreversible actions. In practice, this means agent competitiveness is increasingly a systems-engineering problem, not just a model-quality problem.
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- Agent harness memory loss and compaction are still a major usability problem — adityaag · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21