All videos
0:00 / 0:00
ai

Anthropic Doesn't Trust An Agent to Both Write and Cite. You Shouldn't Either.

The AI Automators16 June 2026Watch on YouTube

Part of series

Ep. 1 · Agentic Engineering Hour

View the series

What you'll learn

  • You learn about subtle AI agent failure modes like overextension, conflation, and citation mismatch
  • You see seven verification strategies for checking citations and sources in agentic RAG systems
  • You understand why Anthropic uses a separate citation subagent in its multi-agent research architecture
  • You discover the difference between evidence-first and claim-first verification approaches
  • You learn how NotebookLM applies its own approach to source verification

Frequently asked questions

What is overextension in AI agents?
Overextension is a failure mode where a source supports a narrow claim, but the model states a broader and more confident version. The error becomes more subtle as models improve.
Why does Anthropic use a separate citation subagent?
Anthropic uses a separate citation subagent because an AI agent cannot both write content and verify citations reliably without risk of errors. Responsibility is separated to limit hallucinations and mismatches.
What is the difference between evidence-first and claim-first verification?
Evidence-first starts with the sources and builds claims from them, while claim-first starts with the claim and checks whether sources support it. Both approaches are treated as separate strategies in the video.
How many verification strategies are discussed in the video?
The video covers seven verification strategies for agentic RAG systems, including the NotebookLM approach, semantic verification, Anthropic's citation subagent, evidence-first and claim-first, and source verification harnesses.

Topics

ai agentsragcitationshallucinationsverification

In this video

AnthropicAnthropic

Read next

Description from the channel

👉 Access our Starter Apps & AI Architects course in our community https://www.theaiautomators.com/?utm_source=youtube&utm_medium=video&utm_campaign=tutorial&utm_content=verified-rag 🔗 PRDs for the citations & source grounding feature (GitHub): https://github.com/theaiautomators/claude-code-agentic-rag-series 🔗 Anthropic's multi-agent research system (citation subagent architecture): https://www.anthropic.com/engineering/multi-agent-research-system 🔗 Docling (bounding-box document parsing): https://github.com/docling-project/docling 🔗 Datalab Marker: https://github.com/datalab-to/marker 🔗 Claude Dynamic Workflows: https://claude.com/blog/introducing-dynamic-workflows-in-claude-code Models keep getting better on the benchmarks, but the failure modes that matter for knowledge agents aren't going away. They're getting subtler. Beyond outright hallucination, there's overextension, where a source supports a narrow claim and the model states a broader, more confident version. There's conflation, where two retrieved chunks get fused into a single claim neither supports. And there's citation mismatch, where the cited source is real and retrieved but doesn't back the paragraph it's attached to. As models scale, these errors don't disappear. They get more confidently stated, which makes them harder to catch. Even Anthropic's own guidance tells users to review Claude's cited sources on high-stakes outputs. If you're building an AI product, you can't leave verification to chance. You have to design for it. In this video I walk through seven ways to build verification into an agentic system. The PRDs for the citation and source grounding feature shown in this video are in the GitHub repo linked above, and the full platform is available in our private community. ⏱️ Timestamps: 00:00 Intro 01:59 1. NotebookLM Approach 06:10 2. Semantic Verification 09:24 3. Anthropic's Citation Agent 11:36 4 & 5: Evidence-First & Claim-First 12:59 6 & 7: Source Verification Harnesses #AI #AIAgents #RAG #Citations #Hallucinations #VerificationLayer #KnowledgeAgents #ContextEngineering #HarnessEngineering #AIBuilder #Docling #NotebookLM #DeepResearch #AIArchitects #GroundedAI