# Marcus Theriault, Staff Engineer at Grovepoint (B2B SaaS, ~90 engineers) — read of SearchAgent API, June 18 2026

> 11 years writing backend services, last 18 months building internal AI tooling. We use Claude API, Cursor company-wide, and I own the agent infrastructure layer. I have a 9-year-old daughter whose soccer warmups are where I catch up on my RSS backlog.

## How I got here

We got burned two weeks ago. Our internal coding agent pulled FastAPI docs from 2022, wrote a middleware pattern against a deprecated hook, and the PR made it through one reviewer before our lead caught it. I searched "search API fresh docs AI agents" and this showed up on page two. Clicked because the meta description matched the exact phrase I'd typed.

## What I clicked first

The headline: "Stop Your AI Agent from Hallucinating Against Stale Docs." That's not marketing speak, that's the exact bug I filed last Thursday. I kept reading. "Source-Ranked Results. Official docs, recent GitHub READMEs, and verified examples rank above blogs." That's the thing I want. The "Token Waste" pain point also landed: "Agents get raw links, then waste tokens parsing 20 pages of noise to find 2 lines of signal." Whoever wrote that has watched agent traces before. Fine so far.

## Where I paused

The four-step "How SearchAgent Works" section. Step 4 says "Drop-In Replacement. Works with Claude Code, Cursor, and any framework that calls web search APIs." That's a real claim and I wanted to see what the API actually looks like. What's the endpoint? What does a response object look like? Is it REST? OpenAI-compatible? Do I swap one env var or do I rewrite my tool schema? The page just... stops. No docs link, no sample JSON, no curl example. For something calling itself an API, there is zero API on this page.

## What I distrusted

Then I hit this: "Honest disclosure: we don't have live customers on this idea yet. We shipped the strategy package; you ship the customer conversations."

I had to read that three times. This isn't a product. This is an idea for sale. The pricing confirms it: "Unlock the dossier $5" and "Adopt the build $99-$199" and "Dossier plus the working code starter." I came looking for a search API I could integrate tomorrow. What I found is someone selling me the business plan to build one myself.

The hero said "Start Free Trial." That implies a running service. There is no service. That's not "honest" framing, that's bait. The "Honest" section buried at the bottom after the pricing table is doing a lot of cleanup work for a top-of-page promise the product can't keep.

Also: "$-22,000 Year-1 take-home" is listed as if it's a feature. Negative twenty-two thousand dollars. For an operator looking to build this. That's genuinely a strange thing to front-load on a product page.

## What would convince me

If this were a real API: one working endpoint I can hit with a curl command, even a sandbox one. A sample response showing me structured output versus raw search results side by side. One concrete integration diff showing before/after in a Claude Code CLAUDE.md or tool config.

If this is an idea marketplace (which is what it actually is): be upfront in the first scroll. Don't dress it up as a live product. I'd actually be interested in the dossier if the positioning were "we researched this pain and mapped the build path, here's what we found" rather than a product page for something that doesn't exist yet.

## What I'd ask in an email reply

1. Is there an actual running API, or is everything on this page describing a future product someone (me?) would build? The "Start Free Trial" button and the "no live customers yet" disclosure are contradicting each other.
2. If the working code starter in the $99 tier exists, what search index is it actually querying and how is freshness determined? Is it a wrapper around Bing/Google with reranking, or something proprietary?
3. The "spam filter" that ranks official docs above blogs: is that a static allowlist, a trained classifier, or a heuristic? Because I've seen "official docs" be six versions behind the GitHub README.

## Verdict: on-the-fence

The pain description is accurate enough that I slowed down, but the page is running a confidence game: it looks like a product landing page for a tool I can use today and reveals halfway down that it's actually a kit for building that tool myself. If I'd understood that from the first scroll I'd have a different judgment. As it stands I don't know if I'm annoyed or curious.

---
*Memo by skeptic persona, generated 2026-06-18. Studio breaks own self-grading loop.*