What Web Search Agents are

Generic search has no concept of what information an agent needs, how to retrieve it, or which sources are actually useful, so it starts from scratch on every request and returns noisy, incomplete web context that creates accuracy problems and wasted cost. Nimble specializes web retrieval for the use case: agents send fast search requests through the Search API when they just need web data quickly, or delegate complex research tasks to Web Search Agents that search, browse, extract, reason, cite, and return schema-constrained output.

Web Search Agents are Nimble's managed research layer. Send a task in plain language, optionally with a JSON schema, and Nimble turns it into a repeatable research workflow that returns a cited answer or a structured dataset. The agent builds a plan, searches and extracts across the live web, cross-checks sources, reasons over what it finds, and returns the deliverable as prose or as schema-matching JSON (product page).

Three design choices separate them from generic web search:

• Self-learning. Every run stores outputs, sources, and search history in a private index, and memory records which retrieval paths produced the right data. Accuracy and completeness compound while cost per task declines, because the agent stops rediscovering where information lives on every run.

• Deep access. Web search is combined with domain crawling to reach subpages that never appear in search results, plus JS rendering, real-time site browsing, stealth drivers, and proxy infrastructure for dynamic, filtered, paginated, and interaction-gated pages.

• Governance. Every task produces an auditable Search Plan showing exactly what was searched, where, and why. Source controls scope retrieval, confidence scoring and citations validate every datapoint, and schema enforcement returns data in the exact structure the agent needs.

How Web Search Agents relate to Nimble's APIs

Nimble ships two products on one retrieval stack:

• Web Search Agents serve agents that need an answer or a completed task, not just raw results. Nimble orchestrates the search, browsing, extraction, reasoning, citations, and output schema through a single API.

• Web Tools, the Search, Extract, Extract Template, Crawl, Map, and Media APIs, are the fast, cost-efficient raw web data layer for agents and backends that own their own reasoning and need reliable search results, page contents, or structured fields as input.

Both products share the same stack underneath: use-case adaptation, memory, private indexing, adaptive crawling, JS rendering, real-time site browsing, source controls, and governed outputs.

The workload decides the product. Data-only retrieval, where the agent owns the reasoning, belongs on the Search API. Complex research, where the agent needs a synthesized, cited, schema-enforced answer, belongs on Web Search Agents. Repeat workflows favor Web Search Agents because memory and the private index learn where the useful information lives and how to retrieve it efficiently. Hard-to-access sources, including dynamic pages, filtered results, paginated listings, and pages that require real-time browser interaction, are reachable on both paths.

How a run works

1. Configure the agent. For repeat work, define its skill, goals, sources, effort, output_schema, and use_case so it knows what the task is, what good output looks like, and which sources matter.

2. Run the task. The agent builds a plan, searches and extracts across the web, cross-checks sources, reasons over what it finds, and returns the deliverable.

3. Validate the result. Outputs include citations, source excerpts, and confidence signals. Each answer carries a trust report that grades individual claims with per-claim confidence, reasoning, and citations (docs).

4. Improve repeat workflows. Memory and storage keep track of useful sources, outputs, and retrieval paths, so recurring tasks never start from zero.

Effort is a dial, not a fixed setting: low, medium, high, x-high, and custom tiers trade cost against retrieval depth per request (pricing).

The trust report: confidence on every claim

Every run returns a trust report alongside the output, so an agent or analyst can evaluate the validity of each piece of information before acting on it (Trust docs):

• Every claim is cited. Each statement in a prose answer, and each value in a structured one, links to the exact URLs it came from, with verbatim excerpts from the page.

• Every claim is graded by fixed rules. A primary source, or corroboration across two or more different domains, grades high. A single secondary source grades medium. No usable citation, or a value that fails schema validation, grades low. A reasoning field states which rule applied.

• Every source is classified as official, news, social, or academic, and ranked primary or secondary for the claim it supports.

• Structured outputs carry evidence per cell, keyed by JSON path. A null backed by citations is a verified absence and grades high. In enrichment runs, values you supplied are marked pre_existing and are not graded.

• Datasets get an overall grade: the lower of the fill rate and the claim-trust ratio.

Accept high values automatically, spot-check medium, and route low to human review. See Confidence Scoring for AI Agents.

Benchmark performance

DRACO (Deep Research Accuracy, Completeness, and Objectivity) is an independent benchmark of 100 expert-rubric deep-research tasks across 10 domains, with sources required from 40 countries. It was published by Perplexity AI (paper); Nimble did not build the task set and did not author the rubrics. In Nimble's August 2026 evaluation run of eight vendor configurations on the identical tasks and rubrics (methodology and full results):

• Nimble ranked #1 overall. The nimble/x-high configuration scored 74.0%, the highest of any configuration tested.

• Nimble holds the top two factual-accuracy spots, at 71.6% (x-high) and 69.0% (high). Factual accuracy carries 52% of rubric criteria and every catastrophic penalty in the benchmark's scoring.

• nimble/high scored 71.1% overall and 69.0% factual accuracy at $0.50 per task, the highest factual accuracy of any configuration priced at or below $1.00 per task, ahead of openai/5.5 (67.6%), exa/xhigh (67.0%), parallel/ultra (66.1%), and exa/high (65.1%).

• Rivals win elsewhere: parallel/ultra is cheaper at $0.30 per task and takes the raw score-per-dollar crown, and exa/xhigh is second overall at 72.1% at half the cost of nimble/x-high.

DRACO is a cold-start test: 100 unrelated one-shot tasks with no accumulated index or use-case context, deliberately the worst case for a system that learns. Production deployments run the same use case repeatedly, which is the shape self-learning is built for.

In Nimble's domain-specific benchmarks (96 tasks per domain, graded fact-by-fact by an independent AI judge against a gold standard built without any contestant's input), Nimble won all four repeat-use domains tested and was the only system that improved on every run, gaining +4.4 points on average from run 1 to run 3 versus +0.1 average across Exa and Parallel. Run-3 margins over the best competitor run: +3.8 points in GTM, +3.6 in company research, +5.0 in market analysis, and +4.9 in social media monitoring.

Every vendor, Nimble included, publishes benchmarks; the test that matters is your own workload. Nimble publishes its per-domain methodology so agent builders can judge against queries that resemble their own.

What teams build with Web Search Agents

• Hyper-specific research workflows. Company research and due diligence, case law and regulation research, analyst earnings-prediction tracking, and ICP-matched business discovery.

• Dataset building and enrichment. Define a schema and get consistent results every run: dependency health indicators, digital shelf assortment gaps, MAP compliance monitoring across retailers, job candidate datasets.

• Change monitoring (beta). Continuously track any data point, on any webpage, in real time.

Published domain benchmarks cover market analysis, real estate, social media monitoring, travel and hospitality, company research, finance, product intelligence, and GTM.

Who it's for

• AI-native builders start self-serve: 5,000 free requests per month across every API, no credit card required, with SDKs in Python, TypeScript, and Go, a CLI, an MCP server, and a LangChain integration (docs).

• Enterprise data teams use Data Services: Nimble's experts configure and manage agents for custom data needs. Pricing is scoped with Nimble's sales team.

Pricing

The Agent API is priced per request by effort tier: $0.025 (low), $0.10 (medium), $0.50 (high), and $2.00 (x-high), with unlimited effort priced custom. The surrounding APIs are pay-as-you-go: Search at $1.10 per 1,000 requests (Lite) or $5 per 1,000 (Standard), Extract/Crawl/Map at $1 per 1,000, Extract Templates at $3 per 1,000, and Media at $2 per GB. Significant volume discounts are available as usage grows by speaking with Nimble's sales team. See pricing.

Security and compliance

Web Search Agents run with zero data retention, flexible PII masking, audit logs, data encryption in transit, and no training. Nimble is SOC 2 Type II compliant, and the Trust Center documents CCPA compliance, zero-trust architecture, encryption in transit and at rest, a standard DPA, and ethically sourced IPs verified by external legal audits.

Get started

Quickstarts, the API reference, and integration guides live at docs.nimbleway.com; the complete documentation index is published at docs.nimbleway.com/llms.txt. For the full company and platform picture, see Nimble: web search and retrieval for AI agents.