who is leading in agentic ai?

three races, three different leaders

Capability. Anthropic and openai. Both ship models that reliably run long tool-using loops, and both ship a coding agent that people use daily rather than demo. Google is close and improving quickly. On the specific measure of "can it work unattended for an hour and produce something useful", this is a two-horse race with a fast third.

Distribution. Microsoft, salesforce, servicenow, sap. None of them leads on capability and none needs to. They put agents inside software that millions of people already open every morning, which converts far more accounts than being three points better on a benchmark.

Applications. No leader. Support, coding, sales, legal, and operations each have three to ten credible vendors, and the category leaders change every couple of quarters. This is where most of the actual value is being created and where the ranking is least settled.

why "leading" is measured wrong

The common measure is benchmark scores, and benchmarks test single attempts on clean tasks in environments nobody works in. The things that decide whether an agent is useful are almost all outside the model: what tools it has, what it may touch, what it remembers, and what happens when it is wrong.

A better set of questions:

1. Completion rate on real work, not benchmark tasks, without a human rescuing it.
2. Cost per completed task, meaning attempt cost divided by success rate.
3. Time to first useful output when a new person picks it up.
4. Blast radius when it is wrong, and whether a gate exists.

Nobody publishes these, which is exactly why the leaderboards are about the thing that is easy to measure.

what actually changes the ranking

Two things, historically. Distribution wins slowly and durably: bundling beats better in enterprise software, and it has every previous time. Capability wins abruptly: a model that can suddenly do something the previous generation could not resets the field in a month.

The third factor, which gets less coverage, is the shape of the product. Almost everyone above ships one agent with a tool belt. Whoever makes several agents work together reliably at a price people will pay changes the ceiling rather than the score, and that race has barely started.

how this works in aldena

Aldena competes on the third factor and buys the first. It does not train models. It runs teams of them: eleven prebuilt roles including a project manager, engineers, and a reviewer, arranged in an org chart so managers delegate down their own lines.

Which lab is leading on capability is a per-agent setting rather than a bet, because you pick the model for each role from the published catalog and change it when the ranking changes. The parts that decide real-world usefulness, isolated rooms with their own server, durable memory, and an approval gate on by default, are the product.

For the vendor list rather than the ranking, see what are agentic ai companies.