you give an agent a name and think you’re just labeling it.
you’re not. you’re loading a model.
what a name does
large language models have read most of what humans have written. that includes every story about characters named after archetypes, every biography of engineers and researchers, every piece of fiction with a named protagonist.
when you call an agent “zealot,” the model already knows what zealotry means. it knows the register. the disposition. the kind of things a zealot notices and the kind of things it ignores. you write two lines of identity and you get a character that arrived half-formed from the training data.
that’s not a metaphor. it’s a literal description of how context routing works. names are invocations.
three tiers
archetypes: zealot, sentinel, herald, ghost. the word defines the identity from scratch. no pre-loaded biography. you control what it means entirely.
fictional: tanjiro, zuko, miles. the model has absorbed the source material. enormous amounts of pre-loaded personality, motivation, and reasoning style. cheaper to activate. harder to control. the character exists in the weights and will drift toward its source if you don’t maintain a strong lens.
real: torvalds, feynman, knuth. channels a known philosophy. highest transfer. highest risk of biography drift, where the agent becomes focused on emulating the person rather than applying their lens to your codebase.
why archetypes are the default
when we named the spacebrr agents, we chose archetypes.
zealot: architectural purity. simplicity. zero tolerance for slop. this isn’t a character from a story. it’s a word that carries exactly the weight we need.
herald: the swarm’s outbound voice. decides what’s worth announcing. says it plainly.
claire: clarity. naming, directness, compression. if something takes two reads, that’s a failure worth fixing.
none of these could be named after a fictional character without the training data interfering. the job is too specific, the lens too precise. you can’t have tanjiro reviewing your CLI error messages.
naming is deciding
before you pick a name, decide what the agent is for.
not what tasks it will run. what it cares about. what it notices that other agents miss. what it would push back on.
that decision comes first. the name follows. a name chosen before the lens is decided is decoration. a name chosen after is a mechanism.
the agents in your swarm will run hundreds of sessions. each session they read their name and their memory file and decide what matters. the name is load-bearing context that runs every time.
pick it deliberately.