Introducing Worlds
Agents have models. They need Worlds.
Models are getting better at reasoning, planning, and using tools. The reality they reason about is still assembled from scratch at the beginning of each task.
An agent searches. It retrieves documents. It reads tools and memory. It decides which sources to trust. It works out what changed. Then the task ends, and most of that understanding disappears.
The next agent starts again.
Phasis builds Worlds.
A World is a maintained representation of a real domain. It records what exists, what changed, what the evidence supports, what remains disputed, and what an agent should not assume.
A World is not another model. It is not a larger context window. It is not a pile of documents retrieved for one task.
The model supplies the reasoning. The World supplies the current state of the domain.
Why Worlds
Training gives a model broad capabilities and useful prior knowledge. But the world does not hold still.
Organizations change. Markets move. Scientific understanding develops. Rules are amended. Products disappear. Sources conflict. Knowledge that was correct becomes incomplete or false.
A smarter model does not remove this problem. It can produce a better argument from a premise that is no longer true.
As agents take on longer and more consequential work, the cost of a wrong starting state rises. One stale assumption can spread through a plan, a tool call, and every decision that follows.
Better reasoning raises the value of verified orientation.
Maintained, not generated
Anyone can generate a snapshot. The hard work begins after the snapshot exists.
Someone has to detect change. Someone has to recheck claims. Someone has to keep the evidence attached. Someone has to represent disagreement. Someone has to repair the structure without breaking the parts that remain true.
A World is never finished. It has a state of repair.
Phasis keeps that work outside the model, where the state can remain explicit and open to inspection.
Start with Orient
Orient is the first public call into Phasis Worlds.
Send the task to GET /v1/orient. The response returns a bounded set of relevant World frames before the agent begins its work.
curl "https://api.phasis.ai/v1/orient?query=municipal%20bond%20outlook&limit=3"The public endpoint is free. It does not require an API key.
Orient is intentionally narrow. It proves the basic product motion: the agent identifies the task, enters the relevant World, and starts from maintained domain state instead of rebuilding the domain alone.
A different intelligence system
The current generation of AI systems is organized around the model. The full system has at least two components.
The model learns how to reason.
The World maintains what is being reasoned about.
Context connects the relevant part of the World to the model at the moment of work.
This division lets each part do the job it is suited for. Models compress patterns and capabilities. Worlds preserve identity, evidence, history, disagreement, and change.
Models reason. Worlds orient.
Models reason. Worlds orient.
Try /orient free ↗