/ platform · ai & agents
AI is only as useful asthe technology context it can trust.
An agent asked to act on your estate needs to know what exists, what it means and what it is permitted to do. Most AI systems are given prose and asked to infer the rest.
/ two kinds of context
One describes technology. The other is a model of it.
Generic AI context
Documents, prompts, API descriptions and snippets. A description of a system. The agent still has to infer what any of it means.
Inferrex context
Systems, schemas, fields, relationships, versions, provenance, permissions and capabilities. A model of the system. The agent can act on it.
/ context → capability → control
Three things travel together, or none of them is safe.
Context
The agent receives structured understanding: what the system is, what it holds, how it relates to others, which version is running and where that understanding came from.
Capability
With that context it can resolve, compare, transform and act, rather than guess and be corrected.
Control
Permissions, provenance and boundaries travel with the context. What an agent may do is defined by the same model that tells it what exists.
/ a task, end to end
Identify every customer record affected by a CRM schema change.
The agent reads the new specification and the previous understanding. It identifies the changed fields and their canonical keys, then follows those keys to every other comprehended system holding the same concept, and to the automations and integrations that touch them. It returns the affected records, the downstream systems, a proposed repair for each, and the evidence for every claim: the specification version, the field, the relationship and its provenance.
/ on InferrexMCP
A means of access, not the proposition.
InferrexMCP is how an agent reaches the understanding. The value is what is on the other side of it. The docs cover the surface; The Corpus is what it reaches.
An agent needs the schema, not a document. API schema inference is the context. MCP for enterprise data is how the agent reaches it.
Bring an agent that keeps getting your systems wrong.
The failure is almost never the model. It is that nothing told it what the fields mean.