# Research agents that cite every claim and browse only where you allow

> Cited web research agents: every claim carries a source URL, egress allowlists control where agents browse, outbound domains visible per run. Humans approve what circulates.

- Canonical: https://www.brahmalabs.io/use-cases/research/
- Department: Research & Intelligence
- All departments: https://www.brahmalabs.io/use-cases/

## Where research time goes

Research teams pay for every finding twice: once to find it, once to verify it. Generic LLM output makes the second cost worse — an uncited summary has to be re-verified line by line before anyone puts their name on it, which is often slower than doing the research directly. The findings that do survive verification then scatter into decks and shared drives, where they stop compounding; the same question gets researched from zero two quarters later. And the obvious fix — let an agent browse — is exactly what a security team is right to question: unbounded outbound traffic from inside your perimeter, with no record of where the agent went or what it touched.

## What changes with governed agents

Browsing gets boundaries and a paper trail. Research agents run in sandboxed containers behind network egress allowlists — a DNS and firewall sidecar decides which domains resolve, and the run page shows every outbound domain the agent actually contacted. The security review stops being "trust the agent" and becomes "read the list."

Citation becomes structural instead of aspirational. Briefs land as versioned artifacts in which claims carry source URLs, so review means spot-checking links, not reconstructing provenance. Monitoring targets live in a Data Table the schedules read: adding a competitor is adding a row, not editing a workflow.

Verified findings stop evaporating. Approved markdown is chunked, embedded, and indexed in the platform's vector store, and any agent granted access can ground its answers in it through a read-only search_kb tool whose scope is fixed server-side. Retrieval returns passages with their source file and a relevance score — the corpus stays citable, not merely searchable. Long engagements hold together because runs are durable: a document pack that waits three days on a reviewer resumes exactly where it parked, without burning compute in the meantime.

## What stays human

Judgment about significance. An agent can report that a competitor changed pricing; whether that warrants a strategy response is a human call. Curation is a human gate too — search_kb is only as trustworthy as its corpus, so nothing enters the knowledge base without a reviewer's approval. And distribution: no brief, delta, or diligence pack circulates until a named person signs off on it. Agents propose; humans decide.

## The workflows

### 1. Cited web research on demand

- **Trigger:** A researcher submits a question through a form, or messages the research agent in Slack.
- **Flow:** An agent browses inside an egress-allowlisted sandbox, extracts findings, and attaches a source URL to every claim; a verification pass checks that each citation actually supports its sentence.
- **Human gate:** The requesting researcher reviews the brief in the approval inbox; claims without a working source are grounds for rejection back to the agent.
- **Output:** A versioned brief artifact with inline citations, plus the run page listing every domain the agent actually reached.

### 2. Recurring competitor and market monitoring

- **Trigger:** A weekly schedule fires per tracked competitor, with an overlap policy so a slow run never doubles up.
- **Flow:** The agent revisits allowlisted sources, diffs findings against the previous brief, and drafts a delta summary with a source URL on every change it reports.
- **Human gate:** An analyst approves each delta brief before it circulates; quiet weeks produce a no-change note the analyst can resolve in seconds.
- **Output:** A versioned artifact per competitor — each week appends a version, so the trail of what changed and when is diffable.

### 3. Knowledge curation into the vector KB

- **Trigger:** An approved brief lands, or a curator submits markdown findings for indexing.
- **Flow:** The workflow chunks the markdown heading-aware with overlap, embeds it, and indexes it into the platform's vector store as a crash-safe durable job.
- **Human gate:** A curator approves what enters the KB; only reviewed findings are indexed, so search_kb answers stay grounded in vetted material.
- **Output:** Indexed passages the whole org's agents can query via the read-only search_kb tool, each hit carrying its source file and relevance score.

### 4. Due-diligence document packs

- **Trigger:** A deal lead submits the target and scope through a form.
- **Flow:** Fork-join parallelism fans out one agent per topic area — corporate, market, public record — each browsing its own allowlist and citing sources; a final step assembles the sections into a pack.
- **Human gate:** Two named reviewers must both approve each section (N-of-M); a rejection returns the section to the agent with the reviewer's notes.
- **Output:** A versioned document-pack artifact with per-section citations and the full outbound-domain record for the engagement.

## Platform pieces this leans on

- Egress-controlled sandboxes
- Per-run outbound-domain capture
- Vector knowledge base + search_kb
- Versioned artifacts
- Schedules with overlap policies
- N-of-M approvals
- Durable runs
- Data Tables

---

Brahmalabs — the AgentOps platform for production AI agents. https://www.brahmalabs.io