→ Fat Agent: the foundation model providers (OpenAI, Anthropic, Google) absorb everything. One model, one surface, one relationship. Tools and specialists get commoditized underneath.

→ Solver Market: small specialized models outperform the generalist on every specific task. They compete in an open marketplace. Value accrues to the routing layer, not any single provider.

Built a visual demo where multi-agent AI workflows are expressed as algebraic equations.

Four operators: → chain (serial) + parallel (aggregate) ÷ filter (constrain) ² iterate (refine)

So instead of 50 lines of orchestration code, you write:

  ((R + A) → W)² ÷ C → E

Research + Analyze in parallel, synthesize, write, refine twice, critique, summarize.

The equation is the program. A runtime solves it. You watch intelligence flow through the topology.

Not a framework—a notation. Curious what HN thinks about treating agent composition as algebra rather than software engineering.

Attempt to a white paper: agentic-algebra.pages.dev

Mini demo (bigger one in exploration): https://github.com/moldandyeast/agentic-algebra

Two AI models are placed on opposite sides of a "glass barrier." They cannot use language—only geometric shapes on a 400×400 canvas. Neither knows what the other is. Neither knows if the other can perceive, let alone understand.

Their task: establish communication from nothing.

Game Theory is for a world that no longer exists. It taught us to optimize for probability, turning play into a predictable game of competition and consumption. This is a thesis for the opposite: reclaiming play as our essential tool to master randomness. It’s about learning to create in the chaos of our new reality, not just win in the old one.

Some thoughts on a side quest. I wanted to figure out how I can break out of my own box. What happens if AI plays more freely with my thoughts. So I created myself an idea aquarium. Thoughts now flow freely, can bond with others, spawn new ones, and sometimes even create design briefs...