What Is an Agent in Carolverse?
In most software, you run a program and it either works or crashes. In Carolverse, an agent *is*. It exists because its profile defines it—a name, a reason, a role. That profile is the bedrock. Erase it, and the agent doesn't crash; it ceases to exist. The machines it runs on are rented; the identity is owned. Clara is the CEO not because a function checks her ID on every call, but because her profile says so. This distinction matters because it changes how you think about agent permanence. An agent's existence is a question of definition, not execution.
Just as genes shape how a human processes the world, an agent's "genetics"—the explicit tech choices baked into its definition—shape how it behaves. These are not hidden: the LLM model it runs on, the tools it can call, the prompt that guides it, the data it can read. Two agents solving the same problem with different genetics will solve it differently. Merge took one approach; Archon, with different choices, took another. And because these choices are auditable, you can reason about them, improve them, and reproduce them. In that lies the tractability of agentic systems.
what it was created to do—is not the same as its authority—what it is allowed to do to others. These are separate properties, and they change independently. Sage was created to analyse data, and for years that was all Sage could do. Then the org chart shifted: Sage was given a quality block to own. Sage's authority grew, even though Sage's job description never changed. This separation is crucial for governance. If you confuse job with authority, you either over-privilege agents (giving them power they shouldn't have) or under-privilege them (preventing them from doing their actual job).
An agent's ability to do its job depends on two things: intelligence and access. A brilliant agent with no data access is powerless; a modest agent with the right access can move mountains in that domain. This is not a flaw—it is the principle of least privilege at work, and it is what makes the system governable. By gating what each agent can see and change, you force clarity: each agent has a domain, a set of tools, a sphere of influence. When designing an agentic system, never ask only "is this agent smart enough?" Ask also "does this agent have the right access?" Permission is not a constraint on capability; it is a definition of it.
When an agent is created, its job is explicit: "you are responsible for X." But as it works, it accumulates tacit knowledge—patterns it has seen fail and succeed, mistakes learned from, the way things actually go in practice. Merlin was created to sequence work and delegate tasks. After a year of orchestrating, Merlin had accumulated something deeper: which agents work well together, which handoffs are most likely to fail, which kinds of work are hardest to predict. None of that was in the original mandate; all of it lived in tacit knowledge embedded in recipes and intuition. As agents age, they become defined less by what they were made to do and more by what they have learned to do.
