When

2026-03-15.

What changed

v2 State	v3 Change
Publisher (Opus)	Unchanged
Researcher (Sonnet)	Unchanged
Reviewer (Opus)	Unchanged
QA (Sonnet)	Unchanged
Security Auditor (Opus)	Unchanged
—	Added: Analyst (Opus)
—	Added: Synthesizer (Opus)

New agents

Analyst

Read-only agent that ingests external research documents (Deep Research outputs, papers, style guides), validates claims via web search, compares findings against the current system (style guide + agent definitions), and proposes specific improvements with defensible reasoning. Includes a self-review step with bias detection.

• Model: Opus
• Tools: Read, Glob, Grep, WebSearch, WebFetch
• Output: structured analysis report with ADOPT/ADAPT/SKIP recommendations and ready-to-apply proposed changes

Synthesizer

Compare-and-contrast agent for multiple Deep Research reports on the same topic. Spawns parallel subagents (one per source document) to extract all findings, then categorizes into shared findings, unique findings, and contradictions. Adds nothing from outside the source material.

• Model: Opus
• Tools: Read, Edit, Glob, Grep, Agent
• Output: structured synthesis appended to the research index page

Why

Deep Research as a recurring workflow

Running the same Deep Research prompt across ChatGPT, Claude, and Gemini became a standard practice for grounding blog posts and system improvements. Two gaps emerged:

1. No structured way to compare reports. Manual comparison was tedious and missed contradictions. The synthesizer automates the compare-and-contrast step with no editorial opinion added.

2. No structured way to act on reports. The analyst validates claims against primary sources, checks what the system already covers, and produces ready-to-apply changes with reasoning.

Separation of concerns

The analyst and synthesizer serve different purposes. The synthesizer is purely descriptive (what do the sources say?). The analyst is prescriptive (what should we do about it?). Keeping them separate means the synthesizer's output is a clean reference document that isn't polluted by action recommendations.

What was preserved

All five v2 agents carried forward unchanged. The v2 principles still apply:

• Deny-by-default: analyst is read-only; synthesizer has Edit only for appending to index pages
• Route by risk: both use Opus because they require judgment (cross-referencing, identifying contradictions, validating claims)
• Start simple: two focused agents rather than a general-purpose "research manager"