Plural

Plural helps you answer a simple question that turns out to be hard: which AI is good at the work you care about?

Public leaderboards test models on someone else's problems. Your work is different. It has its own tools, its own rules, and its own idea of what "done well" means. Plural lets you describe that work once, run any model against it, and see honest, repeatable results.

Why we built it

We think intelligence should be plural. Not one brain with a billion prompts. Different people and companies have different goals, values, and ways of seeing the world, and that is a feature.

If an organization spent twenty years learning how it makes decisions, some of that should become learned behavior in its model. The model should get better at being their model. Getting there starts with three questions:

1. Which models are good at the work we care about?
2. Which model should we use for each task?
3. How do we improve further when none of them is good enough yet?

Plural turns those into three steps: evaluate, route, and train.

How it works, in one minute

Think of it like setting an exam for AI.

flowchart LR
  world["The world<br/>(Environment)"]
  work["The assignment<br/>(Task)"]
  grader["The grader<br/>(Verifier)"]
  contestant["The contestant<br/>(Agent)"]
  result["A score you can trust"]
  world --> work
  grader --> work
  contestant --> work
  work --> result

• You describe the world the AI works in, such as a game of Wordle or a support ticket queue. That is an Environment.
• You write the assignment, such as "guess the word CRANE" or "refund this customer". That is a Task.
• You decide how the work is graded. That is a Verifier.
• You pick the contestants: a model plus its instructions. Each one is an Agent.

Plural runs each Agent on each Task, grades the result, and keeps a full record of what happened: every step the AI took, what it cost, how long it took, and the exact version of everything involved. Group Tasks into a Benchmark and you get a leaderboard for your own work.

Some work isn't done alone. Seat a party of Agents in one Task and they can cooperate, like a team splitting a job, or compete, like players at a poker table. Each one sees only its own side, and each gets its own score.

Evaluate, route, train

Evaluate. Run the models you are considering against your own Tasks and compare them on the same test. Because every input is recorded exactly, you can rerun a result months later and get the same experiment.

Route. Once you trust the scores, send each kind of work to the cheapest model that does it well. Sometimes that is a frontier model. Sometimes it is a small specialist. There is no single best model, only the best one for a goal and a budget. The model catalog lists every model Plural can reach, with prices, and one Plural API key calls any of them.

Train. When the best model available still is not good enough, the same Environments and Tasks become training material. Plural does not replace your training setup. It gives it versioned work, honest scores, and exact records of every attempt.

Where your work lives

Everything you build lives in a project. On your computer, a project is a folder with one subfolder per Environment, Task, Verifier, Agent, and so on. In the web app at pluralintel.com, the same project shows up in the project switcher at the top of the sidebar.

Every run starts from your own machine. Add --record and the run is also recorded in a private hosted project, where your team sees it in the web app. To run Trials in cloud sandboxes instead of on your machine, give an Environment a cloud Runtime. Nothing is ever public unless you choose to publish it to the Hub.

Where to go next

• Quickstart takes you from sign-up to a scored Benchmark in the web app, in about twenty minutes.
• Core concepts explains every word above in five minutes.
• Bring your own work starts from traces or a system you already have.