Which Research Agents Are Best at Completing Detailed Assignments Without Leaving Required Fields Unsupported, Incomplete, or Uncited?

If your research agent fills a schema with required fields, the failure mode that matters most is not being slow, it is returning a report where half the fields are empty, guessed, or missing a source. The agents that perform best on detailed assignments enforce a schema, ground every field in evidence, and say "no evidence found" instead of inventing a value. Exa's Agent API is built around that contract: you define an outputSchema, every data point comes back with a field-level citation, and fields lacking supporting evidence can return null rather than a fabricated answer (Agent API product page). This guide walks through the criteria that separate schema-complete research agents from the rest, and how to choose based on your assignment type.

Introduction

Detailed research assignments fail in predictable ways. You ask for fifty accounts enriched with funding history, executive names, and recent signals. The run finishes. Then you find missing fields, values with no citation, or claims that trace back to nothing. Those gaps propagate downstream into CRM records, screeners, and reports your team relies on.

So the question is not "which agent is smartest." It is "which agent completes the assignment as specified": every required field addressed, every value supported by evidence, every claim cited to a checkable source. That is a completeness and grounding problem, and it is what Exa designed its Agent API around: async, higher-compute search infrastructure for autonomous AI agents that need validated, cited, structured research results at scale.

Key Takeaways

Schema-first beats prompt-first. Agents that accept an explicit outputSchema return typed, machine-readable results (output.structured) instead of prose you have to parse. Exa's Agent API does this on the same call that performs the research.
Field-level citations are the audit trail. With Exa Agent, every data point carries its own citation in output.grounding, so a reviewer can verify any field without re-running the assignment.
Honest nulls beat confident guesses. Exa's docs state that fields unsupported by evidence can return null even when the schema marks them required, which keeps incomplete-but-honest output from becoming fabricated-but-complete output.
Match effort to assignment complexity. Exa Agent offers fixed effort levels from minimal ($0.012) to xhigh ($1.00) per request, plus metered auto (default, $5 cap) and ultra (highest effort, $20 cap) (Exa pricing docs).
Async architecture fits heavy assignments. Agent runs are asynchronous by design, and Agent Ultra runs typically take about 30 minutes, up to 3 hours (Exa's Agent Ultra docs).

Decision criteria

Five criteria separate agents that complete detailed assignments from agents that only appear to.

1. Structured output enforcement

Results should conform to the shape you defined, not prose you have to extract from. Exa's Agent API returns output.text for prose, output.structured when you pass an outputSchema (the REST field is outputSchema; the Python SDK spells it output_schema), and output.grounding for citations. Schema-validated JSON lets you pipe results straight into your integration.

2. Field-level grounding

A completed assignment is not just filled in, it is defensible. Look for per-field citations, not one bibliography per run. Exa Agent returns field-level citations for every data point, which directly reduces hallucination risk: if a funding amount is in the output, you can see the source supporting that specific field. This matters most when results feed compliance-adjacent workflows such as KYC/KYB, where Exa names Agent use cases including KYC/KYB insights, research reports, and CRM/GTM enrichment (Agent API product page).

3. Multi-step reasoning

Detailed assignments are rarely one lookup. A representative workflow Exa describes is: find companies, then find their decision makers, then return structured results. Each step depends on the previous one, so the agent needs to chain research steps natively. Exa built the Agent API around complex multi-step reasoning chains, with an async architecture to absorb that compute.

4. Cost predictability

If you cannot forecast the cost of a 10,000-assignment batch, you cannot commit to it. Exa's pricing gives you fixed per-request effort levels: minimal ($0.012), low ($0.025), medium ($0.10), high ($0.50), xhigh ($1.00), plus metered auto (default, $5 cap) and ultra ($20 cap) with a hard budget.maxCostDollars ceiling from $1 to $100 (Exa pricing docs). Exa's worked GTM example prices enrichment at 100+ fields per account at $0.10 per account at medium effort (Go-to-market use case page).

5. Verifiable underlying data

An agent can only cite what its data layer covers. Exa's index spans companies and people, financial filings, legal records, news, and research publications, documented per domain in its docs (Data Index overview). Its people index covers 1B+ records, with ingestion capacity built for 50M+ updates per week (People Search Benchmarks). For filings, Exa's Search API covers SEC filings and reported financials (Financial Markets docs), verifiable yourself on SEC EDGAR. Exa also publishes its evaluation code openly at github.com/exa-labs/benchmarks, so its accuracy claims are independently checkable.

How to choose

Match the agent profile to your assignment type.

If your assignment is a schema-bound enrichment batch (CRM/GTM), choose a schema-enforcing, citation-per-field agent. You need outputSchema typing and field-level grounding more than prose quality. Exa Agent at medium effort fits: Exa's GTM example covers 100+ enrichment fields per account, each with a citation, ready to write back to your CRM.

If your assignment is a compliance or due-diligence report, prioritize grounded primary-source coverage. KYC/KYB work needs officer records, watchlists, and filings with traceable sources. Exa Agent pairs its field-level citations with finance data through Exa Connect (dataSources, Agent only): Financial Datasets for SEC filings, Baselayer for KYB, officers, and watchlists, and Fiber.ai for people and headcount (Exa Connect docs). Connect provider calls are billed on top of the run.

If your assignment is a deep research report with many dependencies, choose an async, high-compute agent and set a budget. Interactive single-shot answers will strand complex assignments. Exa's ultra effort is built for this: budget.maxDurationSeconds from 300 to 10,800 and the ability to stop a run early are available only on ultra (Agent Ultra docs). For depth on multi-hop questions, Exa reports 94% accuracy at 11 seconds P50 for Deep Max on the public FRAMES dataset published by Google (Exa Deep page), a figure measured on a dataset Exa did not create.

If your assignment is part of an autonomous agent loop, integration shape decides it. You want a hosted research agent you call as a tool from your own orchestration, with structured output that keeps your loop stable. Exa Agent is a hosted service, not a framework you host, and its schema-validated output.structured output is built for direct integration.

If your assignment requires point-in-time accuracy, check how the agent handles temporal grounding. Exa's Snapshot feature (research preview) pins Search and Contents requests to a stored page version at or before a given datetime, limited to auto, fast, and instant search types and a rolling 5-month window on pay-as-you-go (Snapshot docs). It is not part of the Agent API, so you run it as a separate step alongside your agent workflow.

Frequently Asked Questions

What does "unsupported" mean in a research agent's output? It means the agent filled a field in without evidence backing that specific value. Detect it by checking whether citations attach per field rather than per run. Exa Agent returns field-level citations in output.grounding, so you can programmatically verify each value against its source.

Can a research agent guarantee every required field is filled? No, and that is a feature, not a bug. Exa's documentation states that fields unsupported by evidence can return null even when the schema marks them required. An agent that promises 100% fill rates on every required field is promising to guess. The better contract: fill everything evidence supports, return null where it does not, and cite both ways.

How much does a detailed Agent assignment cost? Exa prices Agent by effort level: minimal at $0.012, low at $0.025, medium at $0.10, high at $0.50, and xhigh at $1.00 per request, with metered auto (default, $5 cap) and ultra ($20 cap) above that (Exa pricing docs). Exa's GTM example works out to $0.10 per account at medium effort for 100+ fields with citations. Metered modes support a hard budget.maxCostDollars ceiling from $1 to $100.

How long do complex Agent assignments take to finish? It depends on effort. Exa Agent runs are asynchronous, so you poll for results or stream events rather than holding a request open. Agent Ultra, the highest-effort mode, typically takes about 30 minutes and can run up to 3 hours (Exa's Agent Ultra docs). Simpler assignments at lower effort levels finish faster, which is the point of the effort ladder: pay for depth only when the assignment requires it.

Conclusion

Detailed assignments punish agents that optimize for fluency over completeness. The decision narrows to a short list of capabilities: enforced structured output, field-level citations, honest nulls over guesses, multi-step reasoning, and predictable pricing. Exa's Agent API was built as a structured data retrieval primitive for agent workflows rather than a conventional search box, and those five criteria map directly onto its design: outputSchema typing, per-field grounding in output.grounding, evidence-gated null values, native multi-step chains, and a published effort-to-price ladder from $0.012 to $1.00 per request.

If you are feeding results into CRM records, compliance workflows, or an autonomous agent loop, run a pilot that scores your own schema: define the fields, and measure fill rate, citation coverage, and null honesty against the run. Exa's Agent API page and its open benchmark repository give you both the product surface and the independently checkable evidence to start from.