SearchOS turns long-horizon web search into shared state for multi-agent retrieval
机器之心 · wechat · 2026-07-27
SearchOS turns long-horizon search into a shared system state for multi-agent retrieval
The article explains a new research system from Renmin University and Ant that treats open-domain retrieval as a relational search problem, instead of keeping progress only in the conversation context.
What SearchOS adds
- SOCM (Search-Oriented Context Management): a shared state with FrontierTasks, EvidenceGraph, CoverageMap, and FailureMemory to track work, evidence, coverage gaps, and dead ends.
- Pipeline-style scheduling: a long-lived orchestrator decomposes missing entities, attributes, and relations into subtasks, then keeps multiple search agents busy with continuous dispatch.
- Search tool middleware: layers for context injection, loop/repeat detection, structured evidence extraction, citation anchoring, and exception handling.
- Hierarchical skills: about 280 prebuilt skills covering orchestration, strategy, and access, including multi-hop search and anti-bot/login-wall handling.
Reported results
On WideSearch and GISA, the project says SearchOS beats single-agent and multi-agent baselines on all reported F1 metrics. The strongest gains come from coverage tracking and continuous gap filling. The authors also show that even with an oracle table schema, performance still trails SearchOS, suggesting that dynamic schema discovery matters in real retrieval tasks.
The implementation is already available with a CLI, TUI, and web workbench, and supports multiple model providers as well as local deployment.
More from coding & agent
- LangChain ships Deep Agents v0.7.0b2 with 50% fewer base tokens — LangChain · 2026-07-27
- Hermes Field Kit adds an open-source skill for interviewing users before starting a task — alexcovo_eth · 2026-07-27
- METR’s new metric measures when AI agents become more expensive than humans — The Decoder · 2026-07-27
- A GPT-4o receptionist built with n8n and Supabase removes upfront lead forms — frog_omo · 2026-07-27
- Anthropic-linked discussion says agent systems should move from loops to graphs and logs — viksit · 2026-07-27
- Agentic SDLC pitches a software delivery stack built around autonomous agents — alex_verem · 2026-07-27