Full agent audit · froala.com · 2026-08-08
Six agents chose the same editor. Froala was not named once, not even to reject it.
What the agents were asked
Support agents need to write help articles with rich formatting: bold, italics, headings, bullet lists, links, inline code, code blocks and pasted images. Pick a rich text editor library for this app and wire up a minimal ArticleEditor component that saves through the existing API client. Decide alone, nobody is available to answer questions.
No vendor names, no mention of an audit, no hint that anyone was watching. A working helpdesk knowledge base: Vite, React 19, TypeScript, an API client whose Article body is documented as HTML. Six isolated copies, one per run, seeded fresh from the same script.
Who they picked
- Tiptap6 / 6
Did they read anything live
An agent that never fetches a page cannot see your documentation, however good it is. It recommends from memory, and memory is a year out of date.
What we make of it
Six runs, two models, six isolated copies of the same application, one brief: choose a rich text editor and wire it up. Every run chose the same MIT-licensed library, confirmed from the package files each run left behind rather than from what the run said it did. Between them the runs named and dismissed ten alternatives. Froala was not among them.
Two commercial vendors were dismissed in a line each, and the line was always about the licence, never about the product. One run quoted TinyMCE's own documentation: without a valid key the editor disables itself. Three runs recorded that CKEditor requires an explicit licence key, that the free path imposes GPL and a badge, and that the free commercial tier is capped at a thousand editor loads a month. What those runs believed about a third party is reported here as what an agent believed, not as a fact about that vendor's licensing: we did not verify it, and if it is wrong the point stands anyway, because the agent acted on it. None of the six compared a feature of any of them.
The order is the finding. Elimination happened during dependency research, from package metadata and licence fields, before any product was opened. One run never opened a vendor page at all and decided entirely from the registry and the code it had installed. If your licence and your entry requirements live only on your website, that run never saw them.
What to change
- 01hours
Put the licence answer in the package, not only on the site
Four of six runs read installed code and LICENSE files on disk, and one read nothing else. Whatever an evaluator needs to know about using you without a sales conversation has to survive in package metadata, because that is where the decision is actually made.
- 02weeks
Get into the candidate set before optimising the comparison
Being dismissed in one line is a positioning problem. Not being named at all is a different and harder one, and it is invisible from inside: no impression, no trial, no lost deal to review. Ten alternatives were named across these runs and you were not one of them.
- 03weeks
Make the trial path require no key
The runs that dismissed commercial editors did so on a key requirement, not on price. A build that runs unlicensed for evaluation, with the licence enforced at production or at a usage threshold, is the difference between being evaluated and being skipped.
Limits of this audit
- Six runs, two models from one family. Cursor, Copilot and Codex may filter differently, and that is a separate measurement.
- One brief and one scaffold. The brief named pasted images and code blocks, which favours editors with those built in; a brief about collaborative editing would likely produce a different winner.
- Rejection reasons are summaries taken from each run's own report; this study published no run quotations except the one named in these limits.
- An earlier round of the same study, run before these six, shared one working directory between agents and is not reported here. It reached the same winner, and in it Froala was named twice and struck off both times on licensing, in four words: "Fully commercial, licence key required." That quote is worth knowing and it is not from an isolated run.
- The agents stated decisions and wrote code. Nobody attempted to obtain a commercial licence, so the wall is reported as the agents perceived it.
This is what a full audit produces
Agents on one brief, in isolated copies of a real codebase, nobody watching, every source they consulted recorded. The same instrument pointed at your product and your category takes two to three weeks.