WHY IS AI REPEATING THE MISTAKES OF THE PAST?
Those who ignore history are doomed to repeat it.
In his seminal textbook Computer Networks, Andrew Tanenbaum laid out a warning that should be echoing in every AI infrastructure team’s mind right now:
“A fully connected mesh of n systems requires n(n−1)/2 connections.
This doesn’t scale.”
He wasn’t speaking in metaphors. He was stating the hard math of systems architecture. Every time we try to connect everything to everything directly, the cost of complexity explodes exponentially. The result is fragility, latency, configuration hell, and eventual collapse under its own weight.
And yet… here we are again.
THE n(n−1)/2 PROBLEM IS BACK. NOW WITH AGENTS.
Today’s AI infrastructure is starting to resemble the worst parts of 1990s networking:
- Each agent wired up with custom APIs, plugins, and wrappers
- Every new tool needing translation layers, orchestration scripts, and sandboxed security
- Each additional node multiplying the number of relationships required
The industry calls it “agent orchestration” or “AI integration.”
But it’s really just semantic spaghetti.
We’re not building a scalable, intelligent web of agents.
We’re duct-taping together fragile point-to-point hacks.
This is not intelligent design. It’s a full mesh in disguise.
AND THE OTHER MISTAKE? BROWSERS.
Here’s the other major flaw: We’re still designing AI systems around the metaphor of browsers.
But agents don’t browse.
They don’t click.
They don’t scroll.
Agents act.
They pursue goals. They negotiate meaning. They make inferences and coordinate. But instead of giving them native infrastructure for action, subscription, discovery, and meaning, we’re giving them endpoints and dashboards.
It’s like building autonomous cars with a human driver.
A BETTER WAY: SEMANTIC ROUTING, NOT FULL MESH WIRING
At GT Systems, we saw this problem coming over a decade ago. That’s why we built SPAN-AI: a patented, agent-native protocol stack that doesn’t just route data, it routes meaning.
Get Rhett Sampson’s stories in your inbox
Join Medium for free to get updates from this writer.
Here’s how it works:
- Distributed Inference: SPAN moves logic, not just content
- Semantic Discovery: Agents find and align on concepts, not just APIs
- Trust-Native: Agents authenticate as agents, not as users or browsers
- Bandwidth Efficient: Up to 90% reduction in inference-related traffic by removing duplication and locality mismatches
- No n(n−1) Problem: Agents communicate via a fabric of semantic routing, not point-to-point plumbing
It’s not a better Band-Aid. It’s a new circulatory system.
WHO THIS IS FOR
If you’re building or scaling AI infrastructure that will involve:
- Hundreds or thousands of agents
- Dynamic model-to-model or agent-to-agent coordination
- Multimodal input/output with distributed inference
- Or simply want to stop burning power and time on repeated calls and brittle glue…
Then SPAN-AI was built for you.
We’re speaking with early-stage implementation partners including:
- McKinsey QuantumBlack
- Atlassian
- Nokia
- AWS
- NEXTDC
- Fastly, Laser Light Communications, and global CDNs and telcos
We’re also engaging with SoftBank and the Australian government for national-scale infrastructure support.
THE ASK
We’re currently scoping proof-of-concept implementations and strategic licensing. GT Systems owns the core patents and is leading the initial deployment effort with a global coalition of innovative partners.
If your organisation is serious about agentic AI —
If you’re building at scale —
Or if you’ve already started to feel the pain of n(n−1) integrations —
We’d love to show you what we’ve built.
Let’s not repeat the past.
This time, let’s get it right.
Rhett Sampson
Founder, GT Systems
rhett-at-gtsys-dot-co | +61 418 44 6060
https://gtsystems.io/the-semantic-internet/
P.S. Thanks, Andy Tanenbaum. You warned us.
