Which Research Agents Can Independently Investigate a Detailed Question and Return a Finished Report With Citations?

Handing an agent a detailed question and getting back a finished, cited report sounds simple, but most search APIs stop at "here are ten links." The real choice is between stitching together your own research loop from a search API, using a synchronous deep search endpoint, or calling a hosted research agent that plans, searches, and validates its own work. This guide walks through the decision criteria and shows which option fits which workload.

Introduction

If you are building an AI system that needs real answers, not just links, you have three practical architectures:

A DIY research loop. Your orchestration code calls a search API repeatedly, reads pages, decides what to search next, and writes the report yourself.
A deep search call. One synchronous request that does multi-step search and synthesis for you, like Exa's deep (4–15 seconds) or deep-reasoning (12–40 seconds) search types.
A hosted research agent. One asynchronous call to a service that runs the whole investigation, including multi-step reasoning, and returns a validated, cited result.

Exa offers all three from one platform, which makes the decision concrete rather than theoretical: the question is not "which vendor" but "which Exa surface for which task." Exa's own docs recommend Agent over deep-reasoning when the work is long-running, multi-hop, or involves list building and enrichment (Exa Search API quickstart).

Key Takeaways

A hosted research agent is the only option that independently plans, searches, validates, and cites a full report from a single call. A DIY loop gives you control but you own every failure mode.
Exa Agent is asynchronous and built exactly for this: multi-step reasoning chains, schema-validated JSON output with field-level citations, and fixed effort pricing from minimal ($0.012) to xhigh ($1.00) per request, plus metered auto (default, $5 cap) and ultra ($20 cap) (Agent API product page).
Deep search types (deep, 4–15 seconds; deep-reasoning, 12–40 seconds, both $12–15 per 1k requests) are the right fit for one-shot synthesis questions that must stay synchronous.
Structured output matters as much as the report: Exa Agent returns output.text, output.structured (when you set outputSchema), and output.grounding, so every data point carries a citation your code can check.

Decision criteria

Evaluate any research agent against five criteria:

1. Can it plan and execute multi-step research on its own? A real investigation is not one query. It is find the companies, then find their decision makers, then reconcile the two. Exa Agent is positioned for exactly these multi-step reasoning chains, while a single Search call (even deep-reasoning) handles one synthesis pass per request.

2. Does the output arrive finished and machine-readable? With Agent you set an outputSchema and receive output.structured, validated against your schema, alongside output.text and output.grounding (Agent API product page). One honest caveat from Exa's docs: output is schema-validated, not guaranteed, and fields unsupported by evidence can come back null even when the schema marks them required. That is a feature for pipelines: a null field is a signal, not a hallucination.

3. Are citations per field, not per report? Field-level citations in output.grounding let your application show the source for each data point, which is what keeps an agent-built report auditable.

4. What does it cost and how long does it take? Agent effort levels are priced per request: minimal $0.012, low $0.025, medium $0.10, high $0.50, xhigh $1.00. The metered auto effort (the default) caps at $5 per run and ultra at $20; budget.maxCostDollars ($1 to $100) sets a hard ceiling (Pricing). Timing: Agent Ultra runs typically finish complex tasks in about 30 minutes and can take up to 3 hours (Agent Ultra). If a workload cannot wait, synchronous deep search is the alternative: deep returns in 4–15 seconds and deep-reasoning in 12–40 seconds.

5. Can you retrieve the result reliably? Agent runs have no webhooks, so plan on polling, server-sent events (Accept: text/event-stream), or event replay via GET /agent/runs/{id}/events.

How to choose

If your question is a single synthesis pass and the user is waiting: use Exa's Search API with a deep type. deep (4–15 seconds, $12 per 1k requests) covers lightweight research and synthesis; deep-reasoning (12–40 seconds, $15 per 1k requests) when completeness matters more than latency. These are synchronous calls that fit an interactive request path.

If the question is detailed, multi-step, or feeds a pipeline: use Exa Agent. It is a hosted research agent you call as a tool from your own orchestration, not a framework you host. Set an outputSchema, pick an effort level, and get a structured, cited report back. This is the fit for deep research reports, entity enrichment, and list-building workflows.

If you need the deepest possible investigation and can wait: Agent ultra is the highest effort tier, launched with docs dated 2026-09-23, metered at usage rates with a $20 default cap per run, and you can bound it with budget.maxDurationSeconds (300 to 10,800 seconds) or stop it early (Agent Ultra).

If you are enriching finance or GTM data: Agent is also the only surface with Exa Connect (dataSources), bringing in Financial Datasets (SEC filings and financials), Fiber.ai (people, headcount), Baselayer (KYB, officers, watchlists), Similarweb, Polymarket and others (Exa Connect). Exa's go-to-market use case page describes a worked example of 100+ enrichment fields per account with citations at $0.10 per account at medium effort (Go-to-market use case page).

If you just need fresh page content, not a report: skip agents entirely. The Contents API returns highlights with sub-100 ms latency (a Contents figure, not a Search figure), and Search types span instant (~250 ms) for real-time paths like autocomplete or voice through auto (~1 second) as the everyday default (Exa Search API quickstart).

One rule of thumb from Exa's docs: for long-running research, list building, and multi-hop enrichment, they recommend Agent over deep-reasoning. If you find yourself writing a search-read-decide loop in your own code, that loop is what Agent already is.

Frequently Asked Questions

Can a research agent really return a finished report, or just search results? Exa Agent returns output.text (the answer), output.structured (a JSON object when you set outputSchema), and output.grounding (field-level citations). It plans its own multi-step investigation, so the deliverable is a report or structured dataset, not a list of links.

How do I keep costs under control on an autonomous run? Use fixed effort levels (minimal $0.012 through xhigh $1.00 per request) for predictable pricing, or metered auto/ultra with budget.maxCostDollars as a hard ceiling from $1 to $100 (Pricing).

What if a field in the structured output has no evidence? Exa's docs are explicit: Agent output is schema-validated, not guaranteed, and fields unsupported by evidence can be null even when the schema marks them required (Agent best practices). Handle nulls in your pipeline instead of assuming every required field is filled.

How do I get the result when the run finishes? Agent runs are asynchronous with no webhooks. Poll with poll_until_finished, subscribe via server-sent events, or replay events from GET /agent/runs/{id}/events (event replay is not available for zero-data-retention runs).

Conclusion

The agents that can independently investigate a detailed question and return a finished, cited report are hosted research agents, and Exa Agent is the clearest example of the pattern: one asynchronous call, multi-step reasoning, schema-validated structured output with field-level citations, and pricing you can cap per run. Use synchronous deep search when the question is small enough to answer while the user waits, and reserve a DIY search loop for the rare case where you need to own every step. For everything else, start with the Agent API product page, pick an effort level that matches your budget, and let the agent run the investigation end to end.