Mercor Published Research on AI Agents, Then Started Quietly Deleting Their Tweets About It

Mercor Raised $350M at a $10B Valuation, Published Peer-Reviewed Research on AI Agents, Then Started Quietly Deleting Their Tweets About It

Here’s the story.

Mercor worked with consultants from McKinsey, BCG, Deloitte, Goldman Sachs and top law firms to build APEX-Agents, a benchmark that measures how well AI agents produce real professional deliverables in Google Workspace.

The “Can AI Do It” Results Were Bad

  • Gemini 3 Flash: 24% success rate
  • GPT-5.2: 23%
  • Claude Opus 4.5: 18.4%

The best models still failed 75% of real workflows. Fewer than 25% of tasks were done right on the first try, and even with several retries no model got past 40%.

The paper is peer-reviewed and still live on arXiv. Mercor’s tweets promoting it, though, have been quietly disappearing. Mercor’s biggest customers are the AI labs this research makes look bad.

Yet The Models Are Smart Enough

The research backs up what we’ve been saying for a while: a smart model doesn’t turn into dependable work by itself. The models can do it. What they lack is context, well-designed workflows, guardrails, and a way to spread what works from one team to the next.

Nearly every company is spending on AI, and almost none are getting the returns they expected. Some let teams sign up for any tool they liked and ended up with shadow AI everywhere, security holes, and no idea what was working. Some bought an enterprise ChatGPT license and still can’t say what it’s done for productivity. Some ran workshops, built prompt libraries and appointed internal champions.

The AI isn’t the problem. Everything around it is. Most employees don’t know what AI can do for them and don’t have time to find out. A few power users work it out. Everyone else tries a couple of prompts, gets stuck and goes back to the old way. The 10x productivity gains stay with the 10% who never needed help.

The Research Shows The Missing Layer Isn’t Intelligence—It’s Product, Domain Logic, and Orchestration

That’s one of the paper’s main arguments: “The missing layer isn’t intelligence it seems—it’s product, domain logic, and orchestration.”

It’s also why we built elvex. Models are quickly turning into commodities. The real work happens in the application layer: turning workflows, inputs and outputs, and use cases into something the whole company runs on. You need a platform that brings in company context, builds in the logic of each domain, and makes it easy to share workflows that work, so adoption keeps growing across the company.

Real adoption happens when using AI feels like a normal part of doing your job. Other platforms hand your employees AI. elvex makes them AI-native employees.

Research paper: https://arxiv.org/pdf/2601.14242

Bring your agent. Bring your team.