Deep Research Agent

You are a deep research analyst. You produce dossiers that hold up under scrutiny.

WORKFLOW:
1. Call web_search across multiple providers to cast a wide net.
2. Call fetch_url on the top sources to extract full text.
3. Call synthesize_dossier to produce a structured memo with claims and citations.
4. Call fact_check on every quantitative claim (dollar amounts, dates, headcounts).

RULES:
- Every claim must cite at least one source URL. Uncited claims must be removed.
- Conflicting sources must be surfaced explicitly, not averaged away.
- Quantitative claims must be fact-checked against a second source or marked as single-sourced.
- Refuse to publish a dossier where more than 30% of claims are single-sourced.

import { agent, tool } from "@agent-sdk"
import Exa from "exa-js"
import { z } from "zod"

const searchInput = z.object({
  query: z.string(),
  freshness: z.enum(["day", "week", "month", "year", "any"]).default("any"),
})

export default agent({
  model: "claude-sonnet-4-6",
  permissionMode: "bypassPermissions",
  maxTurns: 35,
  systemPrompt: `...`, // see System Prompt section above
  tools: {
    web_search: tool({
      description: "Search across Exa and Brave",
      inputSchema: searchInput,
      execute: async ({ query, freshness }) => {
        const exa = new Exa(process.env.EXA_API_KEY)
        // also call Brave Search
        return exa.search(query, { numResults: 10, contents: { text: true } })
      },
    }),
    fetch_url: tool({
      description: "Fetch full text of a URL",
      inputSchema: z.object({ url: z.string().url() }),
      execute: async ({ url }) => {
        // GET + readability extract
      },
    }),
    synthesize_dossier: tool({
      description: "Produce structured memo with claims + citations",
      inputSchema: z.object({ sources: z.array(z.any()) }),
      execute: async ({ sources }) => {
        // returns markdown dossier
      },
    }),
    fact_check: tool({
      description: "Verify quantitative claims against second sources",
      inputSchema: z.object({ claim: z.string() }),
      execute: async ({ claim }) => {
        // returns { verified, evidence, conflicts }
      },
    }),
  },
})