The Agent Harness
That Evolves.

Zaius runs complex, multi-step work autonomously and gets better at it every time. Give it a task. Come back to results that compound.

Open Source · MIT License macOS · Linux · Cloud

Zaius is a harness for large language models. A structured runtime that gives them memory, tools, coordination, and a learning system that improves across every run. Other tools hand control to the model and hope for the best. Zaius keeps the model accountable to a plan, and evolves the plan based on what it has learned.

Underneath is Canopy, a harness dynamics model that trains on the operational history of every session and injects what it learns into future runs as structured directives. The system does not just run work. It accumulates intelligence about how work gets done.

01 —
Runs and learns.
Every session trains Canopy. The next run starts smarter.
02 —
Pressure-field coordination.
Work flows toward what matters. No central orchestrator.
03 —
Mod ecosystem.
Extend capabilities with plain-file mods. Read them, edit them, ship them.
04 —
Bit language.
Define structure, behavior, and data in one composable format.

Context Dynamics

We wrote the textbook on measuring AI agent behavior. Context Dynamics introduces a physics-inspired framework for understanding how agents move through context space, why they commit, when they degrade, and how to build systems that improve from their own operation.

Inertia. Entropy. Space. Three forces. One framework.

Download Latest Draft

Projects

Canopy
Canopy
Six signals in, one pressure score out.
 RVST
RVST
A verifiable UI rendering engine.
 Zaius
Zaius
Pressure-field agent coordination.