AI Insights · Oct 22, 2026 · Verified
How the big labs build research agents
OpenAI, Anthropic and Google use a coordinator that sends out parallel search workers, then a separate step that checks citations. None of them use a knowledge graph.
Claim
None of the three frontier labs' research products use a knowledge graph. All use agentic search over raw documents and the web with long context and tool use, then a separate citation-verification step. The lab investment in 2026 is in RL for search agents and MCP tool ecosystems, not graph extraction.
Evidence
- Anthropic (2025-06-13): an Opus lead agent decomposes the query and spawns parallel Sonnet subagents, each with its own context window; results are synthesised, then a CitationAgent post-processes the report. It beat single-agent Opus 4 by 90.2% on an internal eval. Token usage explained 80% of BrowseComp performance variance. Agents use about 4x chat tokens, multi-agent about 15x.
- OpenAI Deep Research: a single RL-trained browsing model in a
plan-act-observe loop, exposed as
o3-deep-researchando4-mini-deep-research(2025-06-26). Since 2026-02-10 it can call any MCP server and restrict search to trusted sites. The current underlying model is reported second-hand. UNVERIFIED. - Gemini Deep Research: an explicit editable plan, parallel versus sequential sub-tasks, hundreds of sites, now callable through the Interactions API with managed agents and background execution. The GA date and underlying model version come from third-party sites. UNVERIFIED.
- PaperQA2 (FutureHouse, 2024-09): agentic search, full-text parsing, reranking and contextual summarisation, citation traversal; matched or beat experts on LitQA2. Its successor Kosmos (2025-11) shares a "structured world model" between a data-analysis agent and a literature agent, runs up to 12 hours and about 1,500 papers per run, with 79.4% of statements judged accurate. The world model is structured shared state, not a classical KG.
- "AI Scientist" systems (Sakana v2, Google co-scientist) are multi-agent with critique and ranking loops; results are modest and contested.
Open questions
- Whether Anthropic's Managed Agents or Google's Interactions API are the right runtime for a long-running literature job versus a self-hosted loop.
Sources
- https://www.anthropic.com/engineering/multi-agent-research-system — blog, Thu Jun 12
- https://openai.com/index/introducing-deep-research/ — blog, Mon Feb 09
- https://gemini.google/overview/deep-research/ — docs, Wed Oct 21
- https://blog.google/innovation-and-ai/technology/developers-tools/interactions-api-general-availability — blog, Sun Jun 21
- https://arxiv.org/abs/2409.13740 — paper, Wed Sep 25
- https://arxiv.org/abs/2511.02824 — paper, Tue Nov 04
AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.