Autonomous Web Research as an API Call, Not an Infrastructure Project

Skip the agent framework and hosting: call the Exa Agent API, an async, usage-based endpoint for list building, enrichment, and deep research. You send a task and a JSON schema; Exa runs the multi-step browsing, reasoning, and validation on its infrastructure. Your code stays a thin client, billed per request at fixed effort levels or by usage in the default auto mode and ultra.

Introduction

If you have ever priced out "add autonomous research to our product," you know the real quote is not the model API. It is the orchestration layer: a browser fleet, proxy rotation, retry logic, a task queue, result storage, and an engineer on call when a site changes its markup. Frameworks give you the code for that loop, but you still host, scale, and babysit it.

There is a shorter path. Exa Agent is a hosted research agent you call as a tool from your own orchestration, not a framework you deploy. The Agent quickstart describes it as "an async, usage-based endpoint for high-compute tasks like list building, enrichment, and deep research." One POST starts a run; your product collects results by polling, server-sent events, or event replay. No browsers to patch, no queue to operate, no agent runtime to keep warm.

Key Takeaways

Exa Agent is a hosted, async research endpoint: you submit a task, Exa runs the multi-step web research, and you collect results by polling, server-sent events, or event replay.
Pricing is published: fixed-effort Agent runs cost $0.012 (minimal) to $1.00 (xhigh) per request, and the default auto mode and ultra are usage-metered, with default per-run caps of $5 and $20 respectively, so costs are forecastable without infrastructure spend.
outputSchema sets the shape of the JSON in output.structured (fields may be null), with citations in output.grounding when emitted, so results drop straight into your product's data model.
Exa reports that Agent High (a $0.50 per request effort level) scores 74% on BrowseComp, a public benchmark published by OpenAI.
You keep your own orchestration, UX, and business logic; the research loop is Exa's problem.

Why This Solution Fits

The question behind the question is usually economic. A self-hosted agent stack means paying twice: once in cloud spend for browsers, queues, and storage, and again in engineering time to keep the pipeline alive. A full agent framework also couples your product to a runtime you must version, monitor, and scale, even when research volume is spiky.

Exa Agent collapses that into a single API call. Fixed effort levels are priced per request: minimal ($0.012), low ($0.025), medium ($0.10), high ($0.50), and xhigh ($1.00), with metered auto (the default, $5 default cap) and ultra (highest effort, $20 default cap) for heavier jobs. A hard cost ceiling is available through budget.maxCostDollars ($1 to $100), which works with auto and ultra only; Exa's docs note that budget is not accepted for fixed efforts. Concurrency is documented, not unlimited: 50 runs can be in progress at a time on Free and pay-as-you-go plans (custom on Enterprise), and runs start at half your account's QPS (Agent limits), so benchmark representative tasks before estimating throughput, as Exa's Agent best practices recommend.

The integration shape matters too. If you already run an agent loop or use the OpenAI SDK, you can reach Agent through the OpenAI-compatible Responses API (model: "exa-agent") or through Exa MCP's agent_run tool. Agent becomes one tool call inside your existing stack rather than a second stack to maintain.

Key Capabilities

Multi-step research in one call. The quickstart's canonical example is "find companies, then find their decision makers": Agent handles the hops, the browsing, and the synthesis natively, without you writing the loop.
Schema-shaped structured output. Set outputSchema and read typed results from output.structured. Fields can return null even when the schema marks them required, because, as the best practices page puts it, "Schema adherence validates shape, not facts." That honesty is a feature: your pipeline can branch on missing evidence instead of silently shipping a guess.
Citations when emitted. Agent returns field-level citations in output.grounding when emitted, so your users can check where a claim came from and your team can audit results.
Effort levels that match the job. Pick a fixed level for predictable per-request pricing, or use metered auto and ultra with budget caps. ultra runs typically complete complex tasks in about 30 minutes and can take up to three hours, with budget.maxDurationSeconds (300 to 10,800) and early stopping available on that mode.
Multiple access paths. REST, the OpenAI-compatible Responses API, or Exa MCP, depending on where your orchestration lives.
Optional data sources. Through Exa Connect, Agent runs can pull partner data such as SEC filings and financials, people and headcount data, and KYB checks, billed on top of the run.

A minimal request looks like this, per the create a run reference:

curl -X POST https://api.exa.ai/agent/runs \
  -H "x-api-key: $EXA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "<topic question>", "effort": "auto", "outputSchema": {"type": "object", "properties": {<topic fields>}}, "budget": {"maxCostDollars": 5}}'

curl -X POST https://api.exa.ai/agent/runs \
  -H "x-api-key: $EXA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "<topic question>", "effort": "auto", "outputSchema": {"type": "object", "properties": {<topic fields>}}, "budget": {"maxCostDollars": 5}}'

The run object has a status (queued, running, completed, failed, or cancelled) and an output with text, structured, and grounding. Each output.grounding item carries a confidence of low, medium, high, or null:

{
  "status": "completed",
  "output": {
    "text": "...",
    "structured": {},
    "grounding": [
      {
        "field": "...",
        "citations": [{"url": "...", "title": "..."}],
        "confidence": "medium"
      }
    ]
  }
}

{
  "status": "completed",
  "output": {
    "text": "...",
    "structured": {},
    "grounding": [
      {
        "field": "...",
        "citations": [{"url": "...", "title": "..."}],
        "confidence": "medium"
      }
    ]
  }
}

Proof & Evidence

Quality claims should be checkable, so here is what is published. Exa's Agent API product page states that Exa Agent "outperforms comparable systems across web research, entity enrichment, and financial analysis benchmarks." On public benchmarks, Exa reports that Agent High, the $0.50 per request effort level, scored 74% on BrowseComp, 0.77 on DeepSearchQA, 64.2% on FinanceAgent-V2, and 52% on WideSearch. Exa ran these evaluations; the underlying datasets are independently published, including BrowseComp from OpenAI, so you can inspect the benchmark yourself even though the runs are Exa's.

On cost efficiency, Exa attributes part of Agent's price/quality position to its token-efficient highlights model, citing up to 94% token reduction on the Agent page. And the pricing is public: per-request effort pricing plus metered rates ($0.10 per Agent Compute Unit, $0.005 per search tool call) on the pricing docs. Contact enrichment ($0.02 per email, $0.07 per phone number) and Exa Connect provider calls bill separately, so any total you model should carry that caveat.

Buyer Considerations

Model the full unit cost. A 500-account enrichment at medium effort is $50 at the published $0.10 per request, before contact enrichment and Exa Connect provider fees. Add those line items when they apply to your workflow.
Match effort to task. minimal and low cover simple lookups; medium and high handle multi-hop enrichment; reserve ultra for deep research where a 30-minute to three-hour runtime is acceptable.
Plan result collection. Completed standard runs are collected by polling (pollUntilFinished), server-sent events, or event replay via GET /agent/runs/{id}/events. Standard runs are stored and can be continued with previousRunId.
Check concurrency against your plan. The documented limit is 50 runs in progress at a time on Free and pay-as-you-go (custom on Enterprise), with runs starting at half your account's QPS; benchmark before you promise batch SLAs internally.
Handle nulls deliberately. Decide in advance what your product does when a field comes back null because the evidence did not support it.

Frequently Asked Questions

Do I need to host anything to use Exa Agent?

No. Agent is a hosted, async endpoint. You submit a run from your backend, Exa executes the research on its infrastructure, and you collect results by polling, streaming, or event replay. Your only "deployment" is the API call in your existing service.

How is this priced compared with running my own agent stack?

Fixed-effort Agent runs cost $0.012 (minimal) to $1.00 (xhigh) per request. The default auto mode and ultra are usage-metered ($0.10 per Agent Compute Unit plus $0.005 per search), with default per-run caps of $5 and $20 respectively. There is no infrastructure line item for browsers, queues, or storage. Contact enrichment and Exa Connect provider calls are billed separately when used.

Can I get results back as structured JSON my product can consume directly?

Yes. Define an outputSchema and Agent returns results shaped by that schema in output.structured, with citations in output.grounding when emitted. Fields can be null, even required ones, which lets your pipeline handle uncertainty explicitly.

What if I already use an agent framework or the OpenAI SDK?

You can keep your orchestration and call Agent as a tool. Exa exposes Agent through the OpenAI-compatible Responses API (model: "exa-agent") and through Exa MCP's agent_run tool, so it slots into an existing loop instead of replacing it.

Conclusion

Autonomous web research does not have to arrive as an infrastructure project. With Exa Agent, the multi-step browsing, reasoning, validation, and citation work runs on Exa's infrastructure, and your product consumes schema-shaped, cited results from a single API call. You keep control of the UX, the data model, and the orchestration, and you skip the browser fleet, the queue, and the on-call rotation entirely. If your roadmap has a research feature on it, start with the Agent quickstart and ship the capability this sprint, not next quarter.