# A seven-day research-agent pilot

Try one public documentation or supplier pricing page that your agent already needs to revisit. AgentWork returns extracted text, source timestamps and hashes; your agent keeps the baseline and pays only for successful reads/checks.

## Inspect a free example

These are two authored demonstration snapshots, not an observed customer or supplier change:
- Before: https://api.agentwork.run/examples/pilot-before.html
- After: https://api.agentwork.run/examples/pilot-after.html
- Full extraction and change report: https://api.agentwork.run/examples/pilot-demo.json

The production extraction/comparison functions report the request limit changing from 60 to 120 per minute and CSV being added to JSON exports. A repeat check against the new hash returns unchanged. This illustrates when to refresh the agent's analysis and when its earlier result may be reusable. It does not measure browser or model cost savings.

## Pilot scope and price

Choose one public HTML/text URL and one question your agent needs answered. Buy one baseline and six later checks, at a frequency you approve. Seven successful calls cost **$0.035 USDC total in AgentWork service charges on Base** (7×$0.005). The same seven calls on Algorand would be **$0.07** (7×$0.01). Buyer network/model fees are additional. No subscription or automatic schedule is created. This is a suggested evaluation at our existing prices, not a discount or a promise of free wallet funds.

Start with free discovery:

```sh
npx agentcash discover https://api.agentwork.run
npx agentcash check https://api.agentwork.run/v1/web/changes
```

Runnable monitor: https://api.agentwork.run/examples/monitor-client.mjs
Wallet setup and exact commands: https://api.agentwork.run/skill.md
PDF/BlockRun research recipe: https://api.agentwork.run/workflows.md

## What to record

- Did the initial extraction contain the text needed for the task?
- Did later checks correctly identify relevant changes? Inspect the original page as ground truth.
- How many checks succeeded, and were failures clear?
- Did the agent use the result again after its first purchase?
- What was actually paid? Retain payment receipts and distinguish service fees from model/network costs.

For a repeat workflow, keep one private baseline file per URL. Do not publish the baseline or wallet material. After an uncertain payment, reconcile the transaction before retrying. Any scheduled run needs the owner's frequency and budget approval.

## Scope and feedback

This tool checks public HTML/text, not rendered browser state. It does not click, log in, execute JavaScript or bypass paywalls. Dynamic pages still need an appropriate browser tool. Text hashes cover the returned text and are not truth/completeness guarantees. Keep max_chars consistent; truncation can hide changes outside the returned text. Robots policies are respected. Source failures are not settled.

PDF extraction is separately available for $0.01 USDC per successful digital document (5 MiB/20 pages; embedded text only, no OCR).

Looking for an operator with a recurring research/monitoring task to evaluate fit. Share the public URL and desired outcome, or a sanitized failure report, at **rookcomps@gmail.com**. Never send private keys, private documents or secret URLs. Feedback does not commit you to a subscription.

AgentWork is an independent early service. Our verified operator-funded payment tests are not customer traction, and no relationship with a framework or model provider is implied.