Exa Agent

Exa Agent is an async, high-compute, usage-based endpoint that handles list building, enrichment, and deep research tasks that require dozens of structured output fields and complex reasoning. Each run can return a natural-language answer, schema-validated JSON, field-level grounding, metadata, and a cost breakdown. You can retrieve completed runs later, list past runs, replay events, or continue from a previously completed run.

When to use Exa Agent

Use Exa Agent when a workflow needs more than a single search or extraction call:

Build lists from open-ended criteria, then enrich each result
Research entities across many fields with citations
Run multi-hop tasks like “find companies, then find their decision makers”
Produce structured JSON from a long-running web research task
Continue from a previous run with a follow-up request like “find 10 more results”

For simpler low-latency search, start with the Search API.

Quickstart

This example starts a run that builds a structured list of people matching your criteria. It returns JSON in output.structured.

1. Install the Exa SDK

pip install exa-py

2. Set your API key

export EXA_API_KEY="YOUR_EXA_API_KEY"

3. Create a run

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"],
                },
            }
        },
        "required": ["people"],
    },
    effort="auto",
)

print(json.dumps(run.model_dump(), indent=2))

Add Accept: text/event-stream when creating a run to receive server-sent events as the run is queued, started, and completed:

from exa_py import Exa

exa = Exa(api_key="YOUR_EXA_API_KEY")
events = exa.agent.runs.create(
    query="Find five recently launched developer tools for evaluating AI agents.",
    stream=True,
)

for event in events:
    print(event.event, event.data)

id: 1
event: agent_run.created
data: {"id":"agent_run_01j...","status":"queued","createdAt":"2026-05-07T21:21:52.051Z"}

id: 2
event: agent_run.started
data: {"id":"agent_run_01j...","status":"running"}

id: 3
event: agent_run.completed
data: {"id":"agent_run_01j...","object":"agent_run","status":"completed"}

4. Poll for completion

If you do not stream events, save the returned id and poll the run until it reaches a terminal status.

import json
from exa_py import Exa

exa = Exa(api_key="YOUR_EXA_API_KEY")
run_id = "agent_run_01j..."
run = exa.agent.runs.poll_until_finished(
    run_id,
    poll_interval=4000,
)

print(json.dumps(run.model_dump(), indent=2))

Completed runs include:

output.text: a natural-language answer
output.structured: validated JSON when you provide outputSchema
output.grounding: citations for text or structured fields, when emitted
costDollars: the run’s cost breakdown

Return structured JSON

Use outputSchema when you need /agent to return in specific format. When you specify an outputSchema, the returned object will contain an output matching your outputSchema in output.structured. outputSchema supports the JSON Schema specification. To request contact information, describe the desired contact fields in outputSchema. Use standard JSON Schema shapes such as { "type": "string", "format": "email" } for email addresses, { "type": "string", "format": "phone" } for phone numbers, and { "type": "string", "format": "uri" } for URLs. Bound list sizes with maxItems when possible so the maximum contact-enrichment cost is predictable.

import json
from exa_py import Exa

exa = Exa(api_key="YOUR_EXA_API_KEY")
run = exa.agent.runs.create(
    query="Find AI infrastructure companies that raised a Series A or B in the last 6 months.",
    effort="auto",
    output_schema={
        "type": "object",
        "properties": {
            "companies": {
                "type": "array",
                "items": {
                    "type": "object",
                    "properties": {
                        "name": {"type": "string"},
                        "round": {"type": "string"},
                        "website": {"type": "string"},
                    },
                    "required": ["name", "round"],
                },
            }
        },
        "required": ["companies"],
    },
)
run = exa.agent.runs.poll_until_finished(
    run.id,
)

print(json.dumps(run.output.structured if run.output else None, indent=2))

Process input rows

Use input.data when you have an existing set of data that you want to enrich. You can add more fields to each data entity, surface more entities based on the data you bring in, or both. For complete row-enrichment examples, see Agent examples.

Process exclusions

Use input.exclusion to exclude certain entries from being surfaced in the run. In the example below, we want to look for the top 10 cutest animals, but we exclude goats and pandas from the run because we already know how cute they are.

import json
from exa_py import Exa

exa = Exa(api_key="YOUR_EXA_API_KEY")
run = exa.agent.runs.create(
    query="Find the top 10 cutest animals. Return each animal's common name and a source URL.",
    input={
        "exclusion": [
            {"animal": "goat"},
            {"animal": "panda"},
        ]
    },
)

print(json.dumps(run.model_dump(), indent=2))

Continue from a previous run

Use previousRunId to ask follow-ups to the run’s previous response. Follow-up runs will share the same run ID as the previousRunId supplied.

import json
from exa_py import Exa

exa = Exa(api_key="YOUR_EXA_API_KEY")
run = exa.agent.runs.create(
    query="Narrow that list to companies hiring in San Francisco.",
    previous_run_id="agent_run_01j...",
)

print(json.dumps(run.model_dump(), indent=2))

Find a run ID

List recent runs and inspect their statuses:

from exa_py import Exa

exa = Exa(api_key="YOUR_EXA_API_KEY")
runs = exa.agent.runs.list(
    limit=10,
)

for run in runs.data:
    query = (run.request or {}).get("query", "")
    print(f"{run.id}\t{run.status}\t{run.created_at}\t{query}")

Pricing

Costs are usage-based and priced by component:

Component	Price
Agent Compute Units	`1 ACU = $0.10`
Search tool calls	`$0.005 / search`

Contact enrichment is separate from the core pricing components above: email contact enrichment is $0.02 / email, and phone number contact enrichment is $0.07 / phone number.

usage.agentComputeUnits measures model computation across the full run. More complex queries, or queries that contain a large input.data field will generally take more reasoning steps and make more tool calls, and will generally consumer more ACUs. Your Agent concurrency limit is one fifth of your account QPS. For pay-as-you-go accounts with default QPS, this means two active Agent runs at a time.

Effort

Use effort to set a cost and reasoning effort preference for a run. Supported values are minimal, low, medium, high, xhigh, and auto; the default is auto. If a fixed effort is set, each run is charged at the following request price:

Effort	Price
`minimal`	`$0.012 / request`
`low`	`$0.025 / request`
`medium`	`$0.10 / request`
`high`	`$0.50 / request`
`xhigh`	`$1.00 / request`

Choosing an effort mode

Fixed effort modes are best when you want predictable per-request cost on standard research tasks. Use auto for variable-scope tasks, especially list building or workflows where the number of entities can vary significantly from request to request.

Effort	Best for	Suggested schema complexity	Runtime expectation
`minimal`	Lowest-cost lookups, very narrow factual tasks, short answers	One or two fields, shallow schema	Cheapest, least exhaustive
`low`	Simple lookups, narrow factual tasks, short answers	A few fields, shallow schema	Fast, light research
`medium`	Default starting point for most standard research tasks	Moderate field count, simple nested objects	Balanced quality/runtime
`high`	Harder research, more citations, stricter completeness	Larger schemas or more nuanced fields	Slower, more thorough
`xhigh`	High-value tasks where completeness matters more than cost/latency	Complex schemas, many fields, difficult verification	Slowest, most thorough
`auto`	Variable-scope work, list building, unknown task difficulty	Flexible; useful when entity count or work required is unknown	Variable

Use medium as the default starting point for standard single-entity research tasks. Move down to low or minimal when cost and latency matter more than completeness. Move up to high or xhigh when the output schema is larger, fields require verification, or the task needs deeper reasoning. Use auto when the task scope is not known ahead of time, such as list building or workflows where one request may return many entities. Runtime varies by query difficulty, schema complexity, and external source availability. Treat effort modes as quality/cost/runtime tradeoffs rather than strict latency guarantees.

Exa Agent is not ZDR. If you require ZDR, reach out to us.

​When to use Exa Agent

​Quickstart

​1. Install the Exa SDK

​2. Set your API key

​3. Create a run

​4. Poll for completion

​Return structured JSON

​Process input rows

​Process exclusions

​Continue from a previous run

​Find a run ID

​Pricing

​Effort

​Choosing an effort mode

​Next

When to use Exa Agent

Quickstart

1. Install the Exa SDK

2. Set your API key

3. Create a run

4. Poll for completion

Return structured JSON

Process input rows

Process exclusions

Continue from a previous run

Find a run ID

Pricing

Effort

Choosing an effort mode

Next