Two Brains: Why Dynamic Model Routing Beats Picking One AI

What if AGI is already here, just scattered across a dozen models, and we keep missing it because we're waiting for one of them to do everything?

I switched models three times.

Not because anything was broken. My coding model had just refactored an entire authentication flow without missing a test. Precise, methodical, not a wasted line. But when I pivoted to editing a client article, the same model suggested revisions so flat I could have ironed a shirt on it. Every suggestion was technically correct and completely lifeless.

I switched to the creative model. The draft had a pulse. Then I needed to debug a hook script, and the creative model started inventing function signatures that don't exist.

Three tasks. Three model switches. I'd been doing this for months before I stopped to wonder why. That's dynamic model routing. Each task goes to the model built for it, instead of running them all through the same one.

Two roads. I'd been pretending they were one.

The Split: No Best Model in 2026

Pluralsight published a breakdown of where models stand in 2026, and the story it tells is simple. There is no best model.¹ Claude leads agentic coding. GPT-5 beat everyone on expert-level knowledge. Gemini? Highest human preference scores. Developers at Faros AI started giving them job titles: "The Auditor" for GPT-5.2, "The Architect" for Opus, "The Workhorse" for Sonnet.² Nobody got called "The Everything."

We spent two years asking which model was best. The answer turned out to be a routing problem, not a ranking one. I carry four model subscriptions because each one has gaps.

The underlying concept even has a name: Mixture of Experts. Specialized subnetworks each handle different types of input, and a routing gate decides who gets the job. The technique dates back to a 1991 paper.³ Thirty-five years old. The marketing budget is new. The real leverage was always the harness, not the model.

Even the vendors know. OpenAI's GPT-5 already routes between sub-models internally, picking a fast responder for simple questions and a deeper reasoner for hard ones. They just don't hand you the steering wheel.

Two Hemispheres: Building a Model Router

So I built the router myself.

Claude Enforcer started as a tool for fighting context drift. But I'd been watching coding models and creative models sharpen in opposite directions. I finally built the thing I kept circling. I called it Hemispheric Model Delegation.⁴

I told it which model thinks like a writer and which one thinks like an engineer. The routing logic sits in the same context layer that keeps skills from drifting. When I start a task, it figures out which brain the work belongs to and asks once if I want to switch. Mid-workflow, if the writing needs research, it sends a focused agent to the analytical model. The answer comes back. The creative model never breaks stride. Two brains, one session.

Creating a skill takes one command. I ran /skill-builder new, added the routing rules, and the audit wired them into checkpoints on its own. The whole project is open source, one line to install.

The research came back from a model the article never met.

IDC predicts that by 2028, 70% of top AI enterprises will use dynamic model routing.⁵ I think the timeline is conservative. Anthropic will ship native model-switching in Claude Code within a year. You'd be fixing a bug one minute and drafting an email the next, and the system would handle the routing. No toggling, no thinking about it. They're already doing it inside the models. The only question is when they expose the controls.

---

References

Footnotes

1. Pluralsight. (2026). "The Best AI Models in 2026: What Model to Pick for Your Use Case." Pluralsight ↩

2. Dunlap, N. (2026). "Best AI Models for Coding in 2026: Real-World Developer Reviews." Faros AI ↩

3. Shivanandhan, M. (2026). "How the Mixture of Experts Architecture Works in AI Models." freeCodeCamp ↩

4. Meetze, F. (2026). Claude Enforcer. GitHub ↩

5. Ward-Dutton, N. (2025). "The Future of AI is Model Routing." IDC ↩