SearchOS: Stopping Search Agents from Looping
antgroup · hf · 2026-07-17
SearchOS: Keeping Search Agents from Going in Circles
This work proposes SearchOS, aiming to solve the problem where web search/information retrieval agents easily lose task progress, search repetitively, and waste budgets during long interactions. The authors remodel open-domain information retrieval as "citation-grounded relational schema completion", allowing agents to advance tasks centered around entities, attributes, and source evidence.
System Design
- SOCM (Search-Oriented Context Management): Externalizes implicit search progress into persistent states, including:
- Frontier Task
- Evidence Graph
- Coverage Map
- Failure Memory
- Pipeline-parallel scheduling: Overlaps sub-agent executions and continuously fills uncovered tasks
- Search Tool Middleware Harness: Intercepts model-tool interactions, records grounded evidence, and intervenes when stuck or budget is exhausted
- Hierarchical skill system: Contains strategy skills and access skills, helping reuse search strategies across runs and avoid repetitive pitfalls
Results
On WideSearch and GISA, SearchOS achieved the best performance among all evaluated single-agent and multi-agent baselines.
Related event: SearchOS: A Multi-Agent Framework to End Search Loops(2 posts)→
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