Engineering decisions, architecture deep-dives, and operator perspectives from the team building XSI LodeStone.
Automation has offered two options, and both ask you to give something up: a macro too brittle to trust, or a bot too autonomous to trust. XSI Apprentice is the third option — and it works the way the old one did.
Agent governance gets treated as a policy question. It's an architecture question — a runtime layer that supervises what an agent intends, plans, and does, on any model or cloud or none at all. Here's why it's mostly missing, and why XSI is publishing ours in the open.
The monolith is the wrong unit of design for a long-running, multi-domain, multi-agent system. The right unit is an organism of interchangeable specialists — and an open standard underneath it is the part most readers will want to interrogate.
Five frameworks shipped in one quarter. They give you tracing and guardrails — not proof a regulator would accept of what an agent did, or who authorized it. That's the spec gap — and the moment to land an open one.
Why the AI model explosion creates more risk than opportunity for ISP operators — and how curated, domain-specific model architecture solves it.
Domain knowledge is the difference between an agent that sounds right and one that is right. Here's what 30 years of broadband operations intelligence, through the ETI partnership, gives an agent that a general model can't.
I built FCC BDC compliance automation at ETI Software Solutions. Here is what this obligation actually is, what it actually costs, and why automating it is harder than most people outside the industry understand.
Everyone compares AI on capability; the bill arrives six months later. Here's the real total cost of ownership — cloud API spend versus dedicated hardware — and where the break-even actually falls for an ISP.
Every AI vendor is pitching mid-market enterprises. Nobody is pitching rural ISPs. That is a mistake — and the Tier 2/3 ISP is actually closer to the ideal first customer for agentic AI than any of the obvious enterprise targets.
Gartner predicts that more than 40 percent of enterprise agentic AI projects will be cancelled by 2027. The reason, in almost every case, will be the same: an agent did something unexpected, an operator lost trust, and the organization decided it was not ready.
A demo impresses; a deployment survives an audit. TM Forum alignment — standardized operations and clear autonomy levels — is what separates an AI that looks production-ready from one an ISP can actually run.
Most AI appliances demo well and die in production. The gap between 'it runs' and 'it's reliable' is where embedded agentic systems live or fail — and it's an architecture problem, not a model problem.
Hallucination isn't a bug you can patch out — it's structural. For an agent running commands on live network gear, a confident wrong answer takes services down. Here's why consensus modeling, not a bigger model, is the fix.
Every business should be asking where its AI data actually goes — and almost none are. For ISPs holding CPNI and BEAD obligations, the answer decides whether agentic AI is an asset or a liability.
A practitioner-first account of building real agentic workflows for Tier 2/3 broadband operators — what works, what doesn't, and why sovereign infrastructure matters more than anyone is saying out loud.