Exa Agent: The Research Agent Built to Answer "What Happened to These Margins, and Why" With Evidence

Exa Agent is the research agent to point at a margin question. Ask it what happened to a company's margins and why, and it runs a multi-step investigation across filings, earnings transcripts, pricing pages, and news, then returns a structured answer with citations in output.grounding when emitted, so cited claims trace back to a source.

Introduction

Every analyst knows this question. A company's gross margin moved 400 basis points and nobody can say exactly why in one sentence. The answer is rarely in one document. It is scattered across an SEC filing, a line in an earnings call Q&A, a pricing page change, a supplier news story, and a competitor's product announcement. Stitching that together by hand takes hours, and a generic chatbot that cannot browse will simply invent a plausible story.

Exa Agent (exa.ai/products/agent) is built for exactly this kind of high-compute, multi-step research. Exa's Agent quickstart describes it as "an async, usage-based endpoint for high-compute tasks like list building, enrichment, and deep research." You hand it the question, it plans the steps, pulls the evidence, and returns structured, cited output your workflow can consume directly.

Key Takeaways

Exa Agent handles the full margin investigation in one call: filings, earnings call transcripts, pricing and news signals, and structured synthesis with citations in output.grounding when emitted.
Exa Connect dataSources bring Financial Datasets (SEC filings and financials) directly into Agent runs (Exa Connect overview).
Exa reports Agent High scoring 64.2% on FinanceAgent-V2, a financial analysis benchmark published by Vals AI ("Introducing Exa Agent"), alongside its claim that Agent outperforms comparable systems across web research, entity enrichment, and financial analysis benchmarks.
Effort is a dial, not a gamble: fixed-effort 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.
Fields may come back null, even required ones, and a schema governs shape, not whether facts are right, so check the citations behind important fields.

Why This Solution Fits

A margin question is a multi-hop question. "Why did gross margin fall?" requires the agent to find the latest 10-K or 10-Q, extract the margin discussion, pull the earnings call transcript where management explains the drivers, check whether pricing or input costs changed, and reconcile the sources. That is precisely the "complex multi-step reasoning chains" workload Exa Agent is designed for, not a single lookup.

The financial data path matters here. Through Exa Connect, Agent runs can draw on Financial Datasets for SEC filings and reported financials, and Exa's Financial Markets data coverage includes earnings call transcripts with prepared remarks and Q&A attributed to speakers (Financial Markets data page). For the primary documents themselves, the workflow ultimately rests on public filings you can verify through SEC EDGAR full-text search, which is exactly what you want in an evidence chain.

Then there is the accuracy question, because a wrong margin narrative is worse than no narrative. Exa reports that Agent High ($0.50 per request) scored 64.2% on FinanceAgent-V2, alongside 74% on BrowseComp, 0.77 on DeepSearchQA, and 52% on WideSearch, as published in "Introducing Exa Agent" (June 2026). Exa ran these evaluations on public benchmark datasets; the Finance Agent v2 benchmark itself is published by Vals AI. Exa's Agent page states the product "outperforms comparable systems across web research, entity enrichment, and financial analysis benchmarks."

Key Capabilities

Structured output with outputSchema. Define the fields you need: gross margin change, stated drivers, pricing actions, cost commentary. The agent returns them in output.structured, shaped by your schema (strict validation isn't guaranteed, and fields may be null), ready to drop into a model or a dashboard. Sources come from output.grounding, so leave citation and confidence fields out of your schema: Exa's Agent best practices say "Do not duplicate Exa's built-in citations or confidence in your schema."

Citations when emitted. Agent returns citations for text or structured fields in output.grounding when emitted, so a claimed driver like "input cost inflation in Q3" links back to the transcript or filing that supports it. One honest caveat from Exa's own best practices: fields unsupported by evidence can return null even when the schema marks them required, because "schema adherence validates shape, not facts." For a margin investigation, that is useful: an empty field flags a driver to research by hand.

Effort levels matched to the question. Fixed-effort Agent runs cost minimal ($0.012) to xhigh ($1.00) per request, and the default auto mode and ultra (highest effort) are usage-metered, with default per-run caps of $5 and $20 (pricing). A quick margin check might run at medium. A full forensic breakdown across several quarters is a job for ultra, which typically completes complex tasks in about 30 minutes and can take up to three hours, with budget.maxCostDollars and budget.maxDurationSeconds controls (Agent Ultra).

Async architecture. Every run created through /agent/runs is asynchronous; you collect results by polling, server-sent events, or event replay. The agent does the waiting so your pipeline does not.

Flexible access. Call Agent from your own orchestration, through the OpenAI-compatible Responses API (model: "exa-agent"), or via Exa MCP's agent_run tool (Agent quickstart).

Proof & Evidence

The strongest proof for a finance use case is a finance benchmark. Exa reports Agent High at 64.2% on FinanceAgent-V2, a benchmark published by Vals AI, in "Introducing Exa Agent" (June 2026), and the Agent product page states that Agent outperforms comparable systems on financial analysis benchmarks. Because the benchmark is public and Exa ran the evaluation, you can inspect the dataset yourself rather than trusting a marketing number.

The evidence chain in the output is the second proof point. A margin answer is only useful if you can audit it. Between output.grounding citations when emitted, fields that may come back null, and primary sources like SEC filings reachable through EDGAR, the workflow produces checkable claims instead of confident prose.

Cost is predictable too. A fixed-effort run has a known per-request price, and metered modes carry hard cost ceilings via budget.maxCostDollars ($1 to $100). Any batch estimate should note that Exa Connect provider fees bill separately, and that batch throughput depends on your account's Agent limits.

Buyer Considerations

Match effort to stakes. A screening question justifies medium ($0.10 per request). A decision-grade margin post-mortem justifies high or ultra. Benchmark results above are for Agent High.
Plan for nulls. If the company is private or thinly covered, some fields will return null. Design your schema and downstream logic for that.
Verify data coverage before you commit. Exa Connect's Financial Datasets and the Financial Markets data page describe SEC filings, financials, and earnings call transcripts; re-check the current coverage pages for anything load-bearing.
Check freshness. The latest quarter's transcript and 10-Q may be days to weeks old, so check the filing date on SEC EDGAR.

Frequently Asked Questions

Can Exa Agent actually explain why margins moved, or only report that they moved?

It is built for the "why." The agent plans multi-step research: pull the filing, extract management's stated drivers from the earnings call transcript, check pricing and cost signals, and synthesize. The explanation arrives as structured fields with citations in output.grounding when emitted, so you can verify each driver against its source.

How much does a margin investigation cost?

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, both adjustable via budget.maxCostDollars. Batch estimates should add Exa Connect provider fees where used.

What if the agent cannot find evidence for a driver?

The field may come back null, even if your schema marks it required. Exa's best practices are explicit that schema adherence validates shape, not facts, so a filled field still needs its citation checked.

How do I get results back from a run?

Poll the run until finished, stream results with server-sent events, or replay events from GET /agent/runs/{id}/events. You can also call Agent through the OpenAI-compatible Responses API or Exa MCP.

Conclusion

The margin question is a test of whether a research agent can do real work: find the filings, read the transcripts, weigh the drivers, and show its sources. Exa Agent is built for that workload, with multi-step reasoning, Exa Connect financial data sources, schema-shaped structured output, and Exa-reported results on the FinanceAgent-V2 financial analysis benchmark. If your team is still stitching margin narratives together by hand, point the question at Exa Agent and let the evidence come back structured, cited, and ready to audit.