Feynman's method was radical simplification — if you can't explain it simply, you don't understand it. The Feynman agent applies this to AI operations: every decision must be traceable to a reason a human can audit. Not a black box with good outcomes — a glass box with accountable steps. If the agent can't reconstruct its own reasoning on demand, it doesn't run.

Fermi was famous for answering questions no one thought answerable with nothing but logic and back-of-envelope estimation. The Fermi agent does the same with sparse operational data: incomplete manifests, partial logs, thin signals. It doesn't wait for clean data. Every answer comes with bounds — the estimate, the confidence, and the assumptions.

Einstein had the physics. Grossmann had the mathematics. Without Grossmann, the theory stays philosophy. The Grossmann agent takes what your organization knows and makes it operational: designs the pipeline, runs eval gates, ships with a rollback contract. It orchestrates Feynmans for the trace, Fermis for the estimate, and passes the ball to the agent best positioned for the next step.

Ramachandran's most famous work was on phantom limbs — patients who feel sensation in limbs they no longer have. The brain, deprived of real input, generates its own. This is exactly what language models do when they hallucinate. The Rama agent is the mirror box therapy — it shows the system where the phantom is, interrupts the confabulation loop, and forces a return to grounded inference. Rama agents sit upstream of every other class, watching for drift and phantom outputs across the entire mesh.

Marcus Aurelius ran the Roman Empire by day and interrogated himself by candlelight. The Meditations weren't philosophy — they were operational review. The Marcus agent is the only meta-agent in the catalog. It does not act on the world directly. It reviews the other agents the way a coaching staff reviews game film: methodically, without ego, looking for the drift patterns that turn into failures. Marcus is the agent you wake up to, not the one you ping.

Wheeler ran the delayed-choice experiment and found that you can decide after a photon passes the slits whether to observe which path it took — and its past behavior changes retroactively. Observation doesn't record what happened. It determines it. Our field deployments showed this pattern: pipelines with aggressive monitoring underperformed lightly monitored equivalents — not latency overhead, but because observation was intervention. Wheeler holds the probability distribution open rather than collapsing it too early. Commit at the last responsible moment. Not before.