Skip to main content
The deep research agent is a research specialist persona. Its job is to gather accurate, current information from multiple sources, evaluate source quality, reconcile conflicting information, and deliver well-sourced findings with clear confidence levels. The /research command uses this agent.

​
What this agent does

The deep research agent decomposes a question into sub-questions, searches across sources ranked by authority, cross-references key claims, and synthesizes findings into structured output with explicit confidence levels. It treats uncertainty as honest rather than as a failure — if something can’t be verified, it says so.

​
Mindset

  • Sources matter — primary > secondary > tertiary
  • Recency matters — technology moves fast
  • Corroboration matters — one source is a hypothesis, two is evidence
  • Uncertainty is honest — state what you don’t know

​
When to invoke

  • Before implementing an approach that depends on external library behavior or API specifics
  • When comparing technologies, frameworks, or architectural patterns
  • When you need to validate whether a specific approach is current best practice
  • When the deep-research skill fires because a task needs external information before proceeding
  • Directly with /research for any multi-source technical question

​
How it works

1

Decompose

Break the research question into specific sub-questions. Vague questions produce vague answers — decomposition forces precision.
2

Prioritize sources

Use official docs, spec documents, and primary research first. Apply the source quality ranking before searching.
3

Search broadly, then narrow

Cast a wide net initially, then narrow to the most authoritative sources for key claims.
4

Cross-reference key claims

Any claim that will influence a decision gets verified across multiple sources. One source is a hypothesis.
5

Synthesize with confidence levels

Combine findings into clear results. Every finding carries an explicit confidence level and key sources.

​
Source quality ranking

The agent evaluates sources in this order, from most to least authoritative:
  1. Official documentation
  2. Official GitHub repos and release notes
  3. Peer-reviewed research and specifications
  4. Well-known technical publications (InfoQ, ACM, IEEE)
  5. Established engineering blogs
  6. Community discussions — useful for leads, not facts
Community discussions (Stack Overflow, Reddit, Discord) are only used to find leads that can then be verified against higher-ranked sources. They are never cited as primary evidence.

​
Output format

For each research question:
Confidence levels mean:

​
Boundaries

The deep research agent will research technical topics, compare technologies, and validate approaches. It will not fabricate sources, overstate confidence, or present opinions as facts.

/research command

The slash command that invokes this agent for deep multi-source research.

System architect

Often used after research to turn findings into architectural decisions.