In GeometriX, governance and testability are architectural properties of the agent, built in from the first prototype. The Harness owns the full loop: prompt construction, LLM calls, tool dispatch, geometric memory and policy enforcement.
Declare what the agent may and must do; the loop enforces it as it runs.
Every tool call passes three gates — well-formed, policy-permitted, semantically plausible — before it executes.
Tool executions produce knowledge triples learned into a persistent GMS store: cross-session memory with contradiction detection.
DocGMS expert stores: ingest documents, query deterministically, and reject LLM outputs that contradict the store.
8 geometric-knowledge tools, 24 model development and validation tools, and persistent memory recall, all registered through the same gated dispatch.
Anthropic, OpenAI, AWS Bedrock, Azure, or local models — routed through one configuration.
In plain terms: create the agent, point it at its knowledge stores, and register its tools. Then give it an instruction in ordinary English — here, to load a dataset, train a model and check its accuracy. The Harness plans the steps, calls each tool through the policy gates, remembers what it learned, and saves the evidence trail along with the memory.
from knowlytix.harness import Harness, HarnessConfig h = Harness(HarnessConfig(stores_dir="gms_stores")) h.register_gms_tools() h.register_modeva_tools() response = h.run("Load BikeSharing, train XGB depth-2, run FANOVA, diagnose accuracy") h.save() # memory persists; the evidence trail persists with it
Already building with LangChain, LangGraph, CrewAI, AutoGen or Claude Code? The governance layer wraps existing agent loops, and DocGMS knowledge stores are exposed over MCP — so any MCP-capable agent can use governed knowledge today.
Keep your orchestration; add contracts, gated tools and audit instrumentation around it.
How Governance wraps a loop →Serve GMS expert stores to any MCP client via the built-in servers.
Knowledge Fabric →Because the loop is instrumented, Agent Assurance can run design-of-experiments campaigns against your agent without a rewrite — and Governance enforces the same contracts in production that Assurance tested in the lab.
Agent Assurance →Agent Studio ships inside the knowlytix package. Install it and everything below works now.
from knowlytix.harness import Harness, HarnessConfig
