Multi-agent research: Claude Code skill for web research

The research skill is a Claude Code skill, adapted by AR Labs from Daniel Miessler's open-source Personal AI Infrastructure, that answers a question by splitting it across several researcher subagents (AI agents started by the main session) that search the web at the same time, then combining what they find into one report with sources. It is for anyone who wants a quick, sourced survey of a question rather than a single search. It is our default for "find out about X": comparisons, current events, citation-hunting for a document, a quick survey of a field. When a question needs rigorous source evaluation instead of breadth, the skill routes it to ResearchBot.

The skill is adapted from the research skill in Daniel Miessler's open-source Personal AI Infrastructure (PAI) project, now published as LifeOS, on which our agent setup was originally built. The three modes of 3, 9 and 24 agents, the timeouts, and the idea of mixing researcher types across providers come from PAI. We have since added the hard agent budget, made the smallest mode the default, replaced the Gemini researcher with a general synthesis researcher, and added the route to ResearchBot.

How it works

The orchestrating session breaks the question into sub-questions, one per agent, and launches every agent in a single message. Each agent then runs three to six of its own searches. Three modes set the fleet size:

Each mode uses an equal mix of three researcher types: one that uses Claude's web search with emphasis on academic depth, a general synthesis researcher, and one that queries Perplexity's Sonar API for citation-grounded answers from a provider outside Anthropic. The Perplexity researcher falls back to ordinary web search if its API is unavailable. The outside provider is there so that agreement between researchers is not just one model agreeing with itself.

Each mode has a hard timeout: two minutes for quick, three for standard, ten for extensive. When it passes, the orchestrator stops waiting and synthesises from whatever has arrived, marking each finding's confidence by how many agents agreed. Raw outputs, the query breakdown and draft notes are kept in a working folder, and the final report is archived with a short README.

The judgement it encodes

The agent count is the budget, and nothing recurses. Decomposition happens in exactly two places: the orchestrator splits the topic into assignments, and each agent splits its assignment into searches. Researcher agents are leaf workers. Their tool lists exclude the tool that starts other agents, so they cannot spawn helpers even if a prompt tells them to. An earlier version relied on instructions alone, and one standard-mode run cascaded five levels deep, researchers launching researchers, until it had started 376 agents. The structural limit replaced the instruction.

Breadth is not rigour. Adding agents buys coverage, not verification. A request that needs depth goes to ResearchBot rather than to a bigger fleet.

Partial results beat waiting. A slow agent does not hold up the report; the synthesis says what was covered.

Recover, do not relaunch. After an interruption, the skill reuses the results it already has and fills named gaps with one or two more agents.

What exists today

The skill is the default route for research requests in our sessions. It is not released. The Gemini researcher, which reached Google's model through its command-line tool, has been retired because that route stopped supporting personal accounts.

Example

From a real quick-mode run, abridged. The question was how developers stay oriented while AI agents rewrite large parts of a codebase; the findings later shaped our Comprehend skill. The three agents each took one angle: empirical studies, tooling, and team workflow. The report summary read:

## Key Findings
- The bottleneck moved from writing to reviewing; the core problem is
  "comprehension debt" (silent loss of the human mental model), not
  code volume.
- Practices sort into four layers: process (spec-driven development,
  comprehension checks at merge), in-repo knowledge (diffable C4
  diagrams, ADRs, AGENTS.md), CI enforcement (fitness functions, the
  only layer that can refuse a change), and triage (AI pull-request
  review, hotspot analysis).
 
## Notes
Gaps: the full DORA 2025 report, primary figures for one vendor study
(the fetch was refused), and large-organisation case studies.
Everything organisational found was prescription, not observation.
Vendor claims were flagged where nothing independent corroborated them.

The "Gaps" section is the part we read first: it says what the fleet did not cover, so the reader knows what not to rely on.