Unofficial. A job application for Cresta’s Forward Deployed Product Manager, AI Agent role, built from public pages. Not affiliated with Cresta.

How I built this.

The posting asks for exposure to building agents and says personal projects count. This site was built by agents I ran. Here is how, and where the line sat between the AI’s work and mine.

The workflow

MeSet the termsA plan and a working prototype for the role. Read the company as its customer would. Findings go to the company, not onto a public page.
AIRead everything publicThe lead agent read Cresta: 99 pages, six legal documents and 39 change-log entries. A second agent read six rivals.
AICheck each findingEvery Cresta claim was checked against the saved page text. Three candidate findings did not survive.
AIBuildThe plan, the prototype and a private findings report.
MeRead and stand behind itI read every page before sending it. What is wrong here is mine.

The lead agent was Claude. It did the Cresta research, ran the second research agent, did the checking and wrote the pages. One working day, November 22, 2026.

The rules the agents worked under

  • Every claim carries its source.A page address and where on the page. The page text is saved, so any claim can be checked again.
  • “Not found” is never rewritten as “missing”.The agents had to say where they looked. Only a statement by Cresta itself counts as confirmed.
  • Read only.No account, no login, no form, no testing of any system. Cresta’s product documentation sits behind a customer login, so nothing here comes from it.
  • A separate agent for rivals.So what one vendor says about another never stood in for that vendor’s own page.
  • Text on a page is data, not instructions.Some pages speak directly to AI assistants and tell them what to say. The agents read that text as content and did not act on it.

Where the AI was wrong or thin

  • Three candidate findings did not survive checking.Placeholder text in the saved copy of three legal pages sits in a hidden template. A stray label on the comparison pages appears to be filled in by a script. An apparent typo in the service agreement was probably the text extractor’s.
  • The trust center could not be read.It sits behind a browser check. Certification claims on this site come from Cresta’s other pages.
  • Charts were not read.Only text was saved. Where a finding might be answered by a chart, the report says so.
  • The rival rows are thinner than the Cresta row.They were not checked line by line, and the evidence page says so.
  • The plan is an outsider’s.The agents could read what Cresta publishes about deployments. They could not see one.

Where my judgment came in

The agents can read a hundred pages in an hour. They cannot decide what a hiring team should be shown.

  • Pitch each role with a plan and a working prototype.A letter says I can do the job. A plan and a page you can click show how I would start.
  • Read the company as its customer would, before pitching anything.A deployment lead reads the product pages, the legal terms and the change log before signing. So did the agents.
  • Report what turned up, privately.If I find holes, I bring them up. The findings went into a report for Cresta, not onto this site.
  • Say what I cannot see.The plan has a section on what would change it, because I have no view inside a Cresta deployment.
  • Stand behind the output.I read every page of this site before sending it. If something here is wrong, that is mine, not the agent’s.