For the complete documentation index, see llms.txt. Markdown variants are available by appending .md to any URL or sending an Accept: text/markdown header. An agent skill is available at /.well-known/agent-skills/site-skill.md.
400
Sponsor

Deep Search

Researches a question, evaluates its own findings, and iterates until the answer is complete and cited.

Output will stream here when you run the agent.

Summary

The Deep Search Agent takes a research question, searches the web, and then critiques its own findings — re-searching to close gaps until the question is fully answered. It returns an answer grounded only in the sources it found, with a citation for every claim. Reach for it whenever a single search isn't enough and you need a verified, sourced answer.

Installation

$ pnpm dlx shadcn@latest add @agentcn/flue/deep-search

Composition

agents/
└── deep-search.ts        # Orchestrator + searcher/evaluator subagents
workflows/
└── deep-search.ts        # Bounded iteration returning a cited answer
tools/
└── search.ts             # Web search via the Bright Data SERP API
skills/
└── research/
    └── SKILL.md          # Iterative research procedure

Customization

  • Search provider. Searches run through the Bright Data SERP API — real-time, unblocked results billed per successful request. Set BRIGHTDATA_API_KEY (and optionally BRIGHTDATA_SERP_ZONE). Edit tools/search.ts to point at a different provider — keep the parameters and return shape.
  • Tune the loop. Adjust maxRounds in workflows/deep-search.ts to trade depth for cost.
  • Reshape the answer. Edit the Answer valibot schema to add fields such as confidence or a source list.
  • Swap the model. Change the model on the orchestrator or subagents.