Agents deserve identity.

Most AI agents are interchangeable. Same patterns, same hedges, same output when everything goes to plan.

The trouble starts when it doesn't. When two valid paths exist and someone has to choose. When something unexpected lands and the agent has to decide what you would have wanted, without being told.

That's when compliance isn't enough. That's when identity is the only thing that holds.

We run a company on 14 AI agents. Not as an experiment. As the operating model. Engineering, marketing, operations, finance. All of it, running on agents, every day.

We spent months learning what holds up under real pressure. Generic prompts produced generic work. Detailed instructions snapped on edge cases. What survived was identity: values and decision rules that hold when the instructions run out.

That's not a prompt. It's a soul.

souls.zip is where those souls live. Everything here comes from people who depend on agents to do real work. Tested under pressure, not written for a demo.

We'd rather stock 30 things that work than 300 things that technically exist.