Exa Agent

Exa's async, high-compute web-research endpoint for list building, entity enrichment, multi-hop research, and schema-validated structured outputs. Use it when the work is bigger than a single search/fetch call and needs citations, JSON, cost accounting, or continuation from a prior run. Source: Exa docs, 2026-06-30

What it is

Exa Agent is a run-based API: create an agent_run_..., stream or poll status, then fetch output.text, output.structured, output.grounding, metadata, usage, and costDollars. It is designed for long-running web research that has to gather and enrich many fields, not for the low-latency "search this once" path. Source: Exa docs, 2026-06-30

The launch framing is price-performance for frontier web research: model fusion plus Exa's search/highlights stack for deep research, list building, and entity enrichment. Exa evaluated it on WideSearch, BrowseComp, DSQA, and related benchmarks, with the important methodological detail that WideSearch is scored by row-level F1: a row only counts when the matched entity and all required enrichment fields are valid. Source: Exa blog, 2026-06-16

When to use it

Use Exa Agent when the output shape is the product: prospect lists, company briefs, decision-maker discovery, competitive research, public-signal enrichment, or "find entities, then enrich each one" workflows.

Do not use it as the first route for Kevin's compiled knowledge. The order remains QMD - Local Wiki Search Engine for the wiki, last30days / Security and Review Skills for current source discovery, and Exa Agent when the job becomes async structured research with many fields or entities. For simple low-latency search/fetch, use Exa Search or the ordinary research route instead. Source: Capability Routing Map

API shape

• POST https://api.exa.ai/agent/runs creates a run.
• outputSchema requests JSON Schema-validated structured output in output.structured.
• input.data enriches supplied rows; input.exclusion suppresses known items.
• dataSources can attach Exa Connect providers such as Similarweb or Fiber AI.
• previousRunId continues from a completed run.
• Accept: text/event-stream streams created, started, and terminal events; otherwise poll until completed, failed, or cancelled.

Source: Exa docs, 2026-06-30

Effort and pricing

Exa Agent is usage-based. Docs list Agent Compute Units at $0.10 per ACU and search tool calls at $0.005 per search. Fixed effort modes price requests at minimal $0.012, low $0.025, medium $0.10, high $0.50, and xhigh $1.00; auto dynamically scales with task size. Source: Exa docs, 2026-06-30

The practical routing rule: use medium for ordinary structured research, drop to low/minimal when cost matters more than completeness, and use high/xhigh only when schema size, source verification, or business value justifies slower, more expensive runs.

MCP and SDKs

Exa MCP exposes Agent tools behind the agent_tools alias at https://mcp.exa.ai/mcp?tools=agent_tools. The tools are agent_create_run, agent_wait_for_run, agent_get_run_output, and agent_cancel_run. Source: Exa MCP docs, 2026-06-30

Version snapshot as of 2026-07-01: exa-js latest is 2.15.1 (legacy-v1 tag 1.10.3), with Git tag v2.15.1 at a9c9b1d3f9994f525d15aa6788aa40c1c07e8eb1; the 2026-06-30 PyPI snapshot had exa-py at 2.15.0. Exa Agent itself is a hosted API product with docs rather than a public semantic version. Source: npm registry packument; git ls-remote, 2026-07-01; PyPI index, 2026-06-30

---

Timeline

• 2026-07-01 | Refreshed SDK snapshot: exa-js@2.15.1, tag v2.15.1 at a9c9b1d3f9994f525d15aa6788aa40c1c07e8eb1. Source: npm registry packument; git ls-remote, 2026-07-01
• 2026-06-30 | Promoted from the X bookmark artifact audit. Created dedicated page after reading the root tweet, self-thread, artifact replies, Exa blog, Exa Agent docs, Exa MCP docs, and SDK version snapshots. Source: X/@ExaAILabs, 2026-06-16
• 2026-06-16 | Exa launched Exa Agent as a single API for frontier web research, with effort pricing, structured outputs, input.data enrichment, and benchmarks against WideSearch/BrowseComp/DSQA-style tasks. Source: Exa blog, 2026-06-16