Exa Launches 'Agent': A Single API for Frontier Web Research

AIDeveloper44 Team

The Evolution of Search to Agentic Research

For AI developers, connecting language models to the web has historically meant stringing together multiple components: search APIs, scraping tools, chunking logic, and complex agentic frameworks. Now, following their recent $250M Series C announcement, Exa has officially released Exa Agent. Launched on June 16, 2026, Exa Agent is billed as a single API to access frontier web research capabilities at a fraction of the cost of piecemeal solutions.

Designed specifically to handle high-compute, asynchronous web tasks, Exa Agent simplifies what used to require highly customized RAG pipelines. It natively handles sub-tasking, reasoning, scraping, and synthesizing data across multiple hops to return perfectly structured JSON. Whether you are building financial analysis tools, go-to-market data enrichment platforms, or deep academic researchers, Exa Agent dynamically orchestrates the compute necessary to get the job done.

Inside Exa Agent: Subagents and Model Fusion

Behind the single API endpoint, Exa Agent operates using a complex web of reasoning loops. When handed a large, ambiguous dataset or an open-ended request, the agent divides the task into subtasks and assigns dedicated subagents to research various domains simultaneously.

A major contributor to the API's cost-efficiency and low latency is its use of model fusion and Exa's proprietary token-efficient highlights model. By dynamically selecting the most cost-effective model for a given sub-task and compressing web context (which Exa claims shows up to a 94% reduction in token usage), developers get frontier-level reasoning without the typical frontier-level API bills.

Benchmarking Performance: WideSearch

To quantify the capabilities of their new agent, the Exa team utilized WideSearch, a benchmark introduced in August 2025 designed to evaluate an agent's ability to aggregate and structure atomic information from across the web. The output format in WideSearch is always a table composed of enriched entities.

Exa adopted a strict Row-F1 metric for these evaluations, noting that traditional cell-level F1 scores were "too permissive" because they rewarded isolated correct values even when the agent failed to attach them to the correct entity. Based on their published Row-F1 vs. Cost charts, Exa Agent outperforms major alternatives—including Perplexity Agent models and Parallel Task models—offering higher accuracy at a substantially lower cost per query.

Building with the Agent API

The Exa Agent API is an async, usage-based endpoint designed with modern AI workflows in mind. It supports Server-Sent Events (SSE) for streaming status updates (queued, running, completed), or standard polling for simpler architectures.

One of the most powerful features for developers is the strict adherence to the output_schema parameter. By passing a JSON Schema, developers can guarantee the exact shape of the output. Here is a brief example of how straightforward it is to build a complex enrichment workflow:

import json
from exa_py import Exa

exa = Exa(api_key="YOUR_EXA_API_KEY")
run = exa.agent.runs.create(
    query="Find engineering leaders at AI infrastructure companies that raised a Series A or B in the last 6 months.",
    output_schema={
        "type": "object",
        "properties": {
            "people": {
                "type": "array",
                "maxItems": 10,
                "items": {
                    "type": "object",
                    "properties": {
                        "name": {"type": "string"},
                        "job_title": {"type": "string"},
                        "linkedin_url": {"type": "string", "format": "uri"}
                    },
                    "required": ["name", "job_title", "linkedin_url"]
                }
            }
        }
    },
    effort="auto"
)

In addition to output_schema, the API provides several advanced input configurations:

• input.data: Bring your own data for row-by-row enrichment. You can feed the agent a list of existing prospects or companies to enrich with current web data.
• input.exclusion: Define entities or sources that the agent should ignore during its research loop.
• previous_run_id: Continue a stateful thread. If you initially ask for 10 companies, you can pass the run ID to request "find 5 more just like the last list" without losing the context of the first session.

Transparent Pricing and 'Effort' Modes

Exa offers two mental models for scaling compute and cost: fixed effort modes and raw resource pricing.

For standard predictability, developers can set the effort parameter to a fixed tier. This is ideal for repeated schemas where you want to lock in your margins:

• minimal ($0.012 / request): Lowest-cost lookups, short factual answers.
• low ($0.025 / request): Simple entity lookups.
• medium ($0.10 / request): The default starting point for typical research.
• high ($0.50 / request): Stricter completeness and complex reasoning.
• xhigh ($1.00 / request): Heavily exhaustive, high-value tasks.

Alternatively, by setting effort="auto", Exa will dynamically scale the resources based on task difficulty. Under the hood, usage is billed via "Agent Compute Units" (1 ACU = $0.10) alongside search tool calls ($0.005/search). Notably, Exa separates contact enrichment fees, charging $0.02 per email and $0.07 per phone number found.

The Takeaway

Exa Agent represents a significant leap forward for autonomous web research. By handling the multi-hop reasoning, dynamic tool selection, and token compression on the server side, Exa allows developers to treat complex web research as a straightforward function call. With strict JSON schema adherence and highly competitive pricing, it is poised to become a staple tool for developers building the next generation of data-driven AI applications.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe