Breaking

Anthropic Launches Deep-Research Workflow in Claude Code to Autonomously Fact-Check and Synthesize Multi-Source Reports

October 20, 2026 at 16:07 EDT

Foundation Models
AI Agents
Research & Papers

Anthropic has released /deep-research, a built-in dynamic workflow for Claude Code that automatically coordinates multiple independent agents to investigate research questions, cross-verify claims against each other, and produce fact-checked reports. The workflow, launched as part of Claude Code v2.1.154 on May 28, 2026, breaks a research question into roughly five distinct search angles, runs parallel web searches, and applies adversarial verification where independent agents attempt to refute each claim before synthesizing a final report.

Anthropic • October 2026

Multi-Agent Fact-Checking Engine

Anthropic's /deep-research workflow deploys 75–105 parallel agents that independently search, cross-verify claims, and filter output through adversarial testing before synthesizing cited reports.

105

Agents (documented max)

Verification stages

2.3×

Cost vs. search

~15

Sources fetched

The 5-Stage Workflow

1. Scope: Break question into ~5 search angles

2. Search: Run parallel web searches across angles

3. Fetch: Retrieve ~15 sources and extract claims

4. Verify: 3 skeptic agents test each claim for refutation

5. Synthesize: Combine surviving claims into cited report

STRENGTHS

Adversarial testing catches errors
Multi-source verification increases confidence
Automatic claim filtering removes unverified facts
Structured output with inline citations

CONSTRAINTS

~2.3× more expensive than semantic search
Consensus ≠ truth (bad sources compound)
High token/context consumption
Best for accuracy-critical questions only

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.