🗡️ Offensively Simple

Where we're going, we don't need handcoded state machines.

The entire AI framework industry is a bet that models can't sequence. That without a workflow engine — a DAG, a planner-executor split, a handoff protocol — the model will flounder. So they build the graph. They build the router. They build the agent topology.

The model already knows how to sequence. What it doesn't know how to do is anything real.

The unlock is not orchestration. The unlock is tools.

A tool is a capability with a boundary. Here is something you can do. Here is the shape. Here is what comes back.

Give a model good tools and sequencing emerges. The model doesn't need a flowchart. It will fearlessly descend into any maze by trial and error and find the route the way rainwater finds the riverbed.

The systems that actually work right now are offensive in their simplicity:

• A loop
• A set of tools
• A system prompt
• A stream

That's it. That's the architecture. In the future, the orchestration layer will be considered quaint — the way a React developer scoffs at the bygone jQuery enthusiast.

The failure mode is not "the model chose the wrong step."

The failure mode is "the tool was bad." It didn't give the model feedback that would help it into the pit of success.

Weak tools break everything. Not because the model can't orchestrate — because there's nothing worth orchestrating. Good tools are load-bearing. They expose capabilities that extend AI. Everything else is geocentrism.

If your agent needs a framework to behave, you're rebuilding something the model already does better.

Build better tools. The framework dissolves. What remains is a loop, a boundary, and a stream of events.

Offensively simple.