How we compare ourselves to 13 agent platforms — and where we lose

Most vendor comparison pages are advertisements wearing a table. The competitor column is filled in from memory or from a sales deck, every row the vendor wins is missing, and the page was last touched two product cycles ago. You can usually tell within ten seconds, and once you have noticed, you stop believing the rest of the page too.

We wanted the comparison to be worth reading, so we built it differently.

What we did

Thirty-two capabilities, thirteen competitors — Kognitos, Beam AI, Relevance AI, Gumloop, Lindy, CrewAI AMP, n8n, Zapier, Make.com, Microsoft Copilot Studio, Salesforce Agentforce 360, Google Gemini Enterprise and ServiceNow AI Agents. Every cell was checked against the vendor's own documentation, changelog, API reference or product page. Not against press coverage, not against an analyst quadrant, and not against what a competitor says about another competitor.

Our own column was checked against our source code. That is why several rows read "partly" where a marketing page would happily say yes. A process definition can be created through our API but there is no screen for it, so it says partly. Learned rules can be listed and retired but you cannot see which run they shaped, so it says partly.

Three rules we imposed on ourselves

"Not verified" is not "no." If we could not confirm a capability from a vendor's own material, the cell says so and stays out of the comparison entirely. It does not become a point in our favour. Absence of evidence about a competitor is a gap in our research, and dressing it up as a gap in their product is how a comparison page becomes something worse than wrong.

Where they win, we say so. Every vendor page has a section titled "Where [vendor] is ahead", above the section where we are level. If one of those rows decides your case, buy theirs — you will be happier and we will not spend six months in a deployment that was never going to work. All thirteen pages have something in that section — between seven and eleven rows each.

The page is generated, not written. Both the matrix and the per-vendor pages render from a single research file, so there is no second copy to forget. When a vendor ships something, one file changes and every page follows.

That last rule earned its keep immediately. Our own column was one day old when we built the pages, and it was already wrong in two places: it said feedback on a finished run did not change later runs, and that we had no control over where inference happens. Both had shipped in the meantime. A hand-written comparison page would have carried those claims for months — understating our own product, which is the funnier direction to be wrong in, but the same failure.

What the research found

The clearest piece of open ground is the plan. Every platform we looked at can build a multi-step agent workflow. Of the thirteen, none lets a person read the plan for a specific run and edit it before that run executes. Some show a plan after the fact. Some let you approve individual actions as they happen, which is a different and more tedious thing. Some let you edit the workflow definition, which changes every future run rather than this one.

Approval gates are also usually opt-in — you get autonomy by default and have to go looking for the brake. Ours is on by default, and the confidence score that decides whether a plan needs a human is computed from the plan's own structure — how many steps have an owner, how many quote something the brief actually asked for — rather than asked of the model that wrote it.

Where we lose is less flattering and just as easy to find on those pages. Scheduled triggers: ten of the thirteen document them, one has them partly, two we could not verify — we have none, and today's answer is "point your own cron at a webhook". Per-action approval: ten have it, we do not. Process versioning with history and rollback: eight have it, we have nothing. SOC 2 Type II: all thirteen, us none. ISO 27001 is more even — six have it, five do not, two we could not verify — but we are in the group without it.

Read it

The full matrix is at /compare, with a page per vendor. Every cell carries the date it was researched. If a row is out of date — ours or theirs — tell us at support@smartstaff.io and we will re-check it against the primary source. It is one file, so the correction lands everywhere at once.