Four Agents, One Nervous System

The substrate#

Four systems live in this archive: a pantry that restocks itself, a procurement desk that rides real card rails, an analyst that sells cited research for USDC, and a DJ that plans beatmatched sets from my own library. On the surface they share nothing. Underneath they run on the same spine — and the convergence was discovered, not designed.

Probabilistic imagination; deterministic execution. Procurement says it outright: the LLM drafts an unsigned mandate, and only policy.classify → validate → sign can turn it into authority. JIM's version is a sourcing gate — regex and arithmetic, no model — that rejects any memo with a number it can't trace to a cited fact. The DJ wraps its only two LLM components in a deterministic Critic that fails plans on harmonic math. Grocery Buddy splits the brains formally: Claude owns language, Temporal owns time and money, and prediction is one line of arithmetic neither touches.

Small models for volume, big models for judgment. Haiku routes intents, parses replies, hints at purchase classes, judges faithfulness. Sonnet appears where messy reality must become clean structure — and in the grocery system, exactly once. Nobody gets Opus without an escalation reason.

Deterministic cores, models at the edges. The auth hot path in procurement is a pure function that decides in under three seconds with no network, no model, no clock. Quiet monitor polls in JIM cost zero inference because a materiality gate speaks first. These aren't guardrails bolted onto an agent — they're the chassis the agent is bolted onto.

One nervous system, zero shared code. The DJ ran on agent-core — shared tracing, tier-routed completions, budgets — until honesty retired it: the DJ turned out to be the substrate's only real consumer, so the two functions it actually used moved in-house. Grocery and procurement reached the same verdict from the other direction: shared design DNA, deliberately zero shared code, so a live product's cadence never drags behind a library's churn.

Approval is a feature#

Every system arrived at a human gate from a different direction, for a different reason.

Grocery Buddy stages the cart and hands over a checkout link — the agent has never placed an order, and its money-live gate is five conditions of evidence, not a toggle. Procurement's high-value tier parks a signed-but-unexecuted mandate for a human, with the issuer's card cap as a second, independent boundary. The DJ's gate is taste-shaped: approve the set as data before any compute renders audio. JIM's gates decide when the system may speak — only when something material changed, never in the second person, never as advice.

↳SHIPProcurement Agentfive planted attacks, all blocked by rules the audit log can cite→

The industry frames approval gates as scaffolding to remove once the agent is good enough. Operating four of them says the frame is backwards. The gate is where trust compounds: every approval teaches the human exactly what the system would have done alone, and that record — not accuracy claims — is what eventually earns autonomy. Autonomy is a verdict, delivered by evals, revocable by evidence.

What transfers#

The checklist that now starts every project, in the order the projects taught it:

Gates before features — procurement shipped money control before its first model call. Instrumentation before the first run — grocery had tracing, a cost ledger, and prediction snapshots wired on day one, which is why a week later it could prove what a run costs. Offline-first — JIM's house rule: everything tested with no key, no wallet, no network; the model is optional luxury. Exactly-once edges anywhere money moves. And one honest question per project, posted on the bench, because a system that can't name what it doesn't know yet isn't being documented — it's being advertised.

↳BENCHJIMTrack 0: prove the live legs before adding surface area→ ↳BENCHDJ Agentthe substrate working in its strangest room→

Four rooms, one nervous system. The fifth room will inherit it on day one.