SearchOS: An Open-Source Multi-Agent Search Framework from Renmin University + Ant, Turning Search State into Infrastructure

Having agents search the web for information is no longer novel. But better model search capability doesn't automatically solve the systems problems of long-horizon tasks—context grows blurrier the longer it gets, material already researched has to be researched again next time, and agents fall into dead loops of repeated queries.

SearchOS, newly open-sourced by Renmin University of China and Ant Group, answers with: turn search state into schedulable infrastructure, rather than stuffing it into conversation history.

Resources

• Code repo: github.com/antins-labs/SearchOS
• Paper: arxiv.org/abs/2607.15257
• Project site: antins-labs.github.io/SearchOS
• YouTube demo playlist: Playlist

What Is SearchOS

The blunter analogy: a traditional search agent is an intern who "researches and forgets"; SearchOS is "a search team with a database."

Its core borrows the design philosophy of relational databases: given a natural-language request, the system first creates a relational search schema—T_m defines the table schema, A_m is the columns to fill in, P_m is the primary key distinguishing each row, and R is the correspondence between tables. The entire search process is continuously filling in and completing this relational table.

Four Core Designs

1. Search-Oriented Context Management (SOCM)

Search state lives in the system, not in the conversation history. Four kinds of continuously evolving shared state are maintained:

• Frontier Tasks: a queue of pending tasks, with priorities and dependencies
• Evidence Graph: an evidence graph recording findings, sources, confidence, and the relations between pieces of evidence
• Coverage Map: the real-time completion state of the relational schema (which cells are filled, which are missing)
• Failure Memory: records of failed queries, inaccessible sources, and missing capabilities

2. Pipeline-Parallel Scheduling

An Orchestrator–sub-agent architecture. A long-lived Orchestrator handles unified planning, scheduling, and convergence, dispatching tasks to multiple short-lived Search Agents, with pipeline-parallel mechanics and continuous dispatch.

3. Search Tool Middleware

Three middleware layers abstract away the dirty work of searching:

• Context Middleware: injects relevant state on demand, controlling context size
• Sensor Middleware: detects loops, repeated queries, and stalls
• Evidence Extraction Middleware: structured extraction, unit normalization, citation anchoring, and evidence storage

4. Hierarchical Search Skill Library

Skills are organized in three tiers—orchestration / strategy / access—with about 280 skills preset in the first open-source release:

• Strategy skills: ranked retrieval, multi-hop search, entity disambiguation
• Access skills: handling anti-scraping, login walls, and dynamic pages

Benchmark Performance

max@3 results on two open information-retrieval benchmarks:

SearchOS leads the baselines on every F1 metric and beats the next-best baseline by 13.4 points on Set F1—the gains come mainly from Coverage Map-driven continuous gap filling.

How to Get Started

SearchOS already ships three work interfaces—pick any one:

• CLI: scripted integration into existing pipelines
• Full-screen TUI: watch schema completion progress, agent task flows, and cell-by-cell evidence live in the terminal
• Web research workbench: graphical operation, with support for exiting mid-run and resuming, or coming back to review later

Multiple model providers are supported, and local deployment works too.

Who It's For

Long-horizon, multi-hop, verification-heavy search tasks are SearchOS's sweet spot:

• Competitive research: compare across product dimensions and fill information gaps
• Complete enumeration of lists / catalogs: guarantee coverage, avoid omissions
• Multi-hop information verification: every conclusion carries citation anchoring
• Consulting / investment research / BD / academia: structured management of long-horizon research tasks

If your work includes a "research a pile of material, then organize it into a structured report" step, SearchOS is worth an afternoon to get running.