Exa Slashes Monitors v2 Web Search Costs by 87% for AI Agents

Exa has released Monitors v2, a rewrite of its scheduled web-search product for agents. A monitor searches on a fixed cadence, tracks previously delivered material, and sends a webhook, an HTTP callback to another service, when Exa identifies new and relevant information.

Each monitor continues to execute on schedule and incur a run charge. Event filtering reduces the duplicate results and downstream LLM calls that often make polling expensive.

Polling compounds the bill

Agents commonly monitor the web with cron jobs that rerun a search every few hours. Each execution costs money, and repeated results can trigger additional model calls for classification, comparison, and summarization. Exa cites commercial-agent limits of three to 15 scheduled tasks per user as one sign of that cost pressure.

Monitors v2 keeps a delivery history for each monitor and evaluates unseen pages before issuing an alert. A page must contain relevant information that is new to that monitor. Scheduled executions remain billable at the published run rate, while the filter targets unnecessary downstream processing.

Three stages before delivery

1. Retrieve. Exa searches its index using a time window aligned with the monitor’s cadence.
2. Judge. Previously seen pages are removed. Remaining pages are checked for relevance, source legitimacy, new information, and whether the reported event falls within the required period.
3. Compose. Qualifying pages become a short summary or JSON matching a supplied outputSchema. The response includes citations and confidence levels for structured fields.

These stages allow an agent to consume a compact event payload instead of comparing every search result with its own history.

The index reaches beyond news

Exa says monitors search its full index, which spans company records, products, podcasts, and general web pages.

The podcast corpus includes full transcripts for more than three million episodes. A monitor can therefore detect a competitor mention inside an episode even when the discussion never appears in an article or social post.

One API call creates a monitor

from exa_py import Exa

exa = Exa()
exa.monitors.create({
  "name": "Competitor mentions",
  "search": {
    "query": "Acme Corp product launches, pricing changes, or podcast mentions"
  },
  "trigger": {
    "type": "interval",
    "period": "1d"
  },
  "webhook": {
    "url": "https://example.com/webhooks/exa",
    "events": ["monitor.run.completed"]
  },
})

The call supplies a search query, a daily interval, and the endpoint that receives completed-run events. The snippet assumes configured Exa credentials and an initialized exa_py client.

When a run produces a qualifying update, Exa sends the new pages, citations, and either a summary or structured output. Developers can attach metadata for routing alerts to the correct customer, workspace, or internal workflow. Request options and payload details appear in the quickstart guide.

Good fits for event alerts

• Prediction-market resolution: collect cited evidence for questions such as how many US colleges formally closed by a specified date.
• Portfolio intelligence: track partnership announcements, executive changes, and financing events across portfolio companies.
• Sales triggers: detect launches, hiring changes, and other account signals soon after publication.
• Adverse-media screening: watch counterparties for lawsuits, regulatory actions, and fraud allegations that affect risk scores.
• Product monitoring: alert on price reductions, restocks, and material listing changes.

Precision claims need a recall test

Exa reports that every update delivered by v2 was relevant during side-by-side tests with v1 on production monitors, while v2 generated fewer alerts. The release post omits the number of monitors, test duration, evaluator criteria, and recall, meaning the share of valid events the system successfully found.

Teams evaluating the service can run known-event tests or shadow existing monitors and measure:

• Relevant alerts as a share of all delivered alerts
• Duplicate alerts for the same underlying event
• Known events missed by the monitor
• Delay between publication and webhook delivery
• Citation accuracy and structured-field validity
• Webhook failure, retry, and replay behavior

Scale in dollars

At the published rate, 1,000 monitors running once per day would generate about 30,000 monthly runs and cost $60 for those runs. That estimate excludes downstream model calls and any other Exa usage, though filtering can reduce those costs by limiting which updates reach an agent.

Monitors v2 is available through the Exa API. Current limits and related charges are listed on the pricing page.