Book VI

The Agent Society

The Muscles of the Species — Specialized Workers Under One Brain

ARI is the brain; agents are the specialized workers. No agent is static (Law XII), no agent works outside the Heart's governance (Law XIX), and every agent's output flows back into the genome.

Chapter OneThe Agent Taxonomy

Chapter TwoThe Canonical Agent Card

Every agent in the society is declared the same way — one canonical card, no exceptions. An agent that is not fully carded may not run.

Chapter ThreeAgent Routing Architecture

ARI does not respond blindly to prompts. It classifies intent, routes work, applies constitutional checks, and orchestrates specialized agents before producing an answer or action. This is the shift from chatbot to intelligence operating system.

At the center of ARI is an orchestrator that receives a request, classifies the task, and routes it to the right specialist or workflow. The system does not answer everything in one pass — it first decides what kind of work exists, and only then decides who should do it. Routing is not a technical detail; it is the decision boundary that defines the whole system. ARI treats the routing policy as the center, not the model — the move from “prompt → response” to “goal → task graph → execution policy.”

The orchestration spine

The five routing layers

Routing runs as a layered stack; each layer narrows the decision before the next begins.

The core routes

ARI runs managed execution lanes rather than a single chatbot stream. Every request lands in one first-class route — chosen by task type, risk, latency, and required capability, never by conversational wording alone.

Chapter FourDynamic Team Assembly

Once a request is routed, ARI automatically assembles the right AI team, then synthesizes their work into one unified recommendation. Agents are chosen by five factors: task type (what work exists), domain (which expertise applies), risk (how consequential), confidence (how certain the routing is), and available context (what memory and data are on hand). The team is the minimal sufficient set — never more agents than the task requires.

Chapter FiveThe Human Decision Layer

The Human Decision Layer is constitutional, not optional — it descends from the Covenant (Book I) and Law XIX, and no agent, route, or optimization may bypass it. ARI clearly separates what machines may do from what only humans decide:

Every recommendation states plainly: "Human decision required: Yes / No." ARI supports the Founder / CVO, Chief Integration & Strategy Officer, COO, CFO, Chief of Staff, division heads, venture founders, researchers, builders, investors, and partners — it replaces none of them.

Chapter SixCore Workflows

Every workflow is declared with the same structure — trigger, agents, outputs, human checkpoints, memory writes, completion criteria — so any workflow can be governed and audited identically. The canonical five:

Chapter SevenAgent Governance

The bridge between taxonomy and operation. Governance is what makes a crowd of agents a society.

Lifecycle

An agent is created only by registering a complete card (Chapter Two); it is activated within its scope, may be paused, and is eventually retired or superseded by a new version — its changelog and history preserved.

Authority boundaries

Each agent's card fixes what it may decide autonomously (low-risk, in-scope, reversible) versus what it must escalate (anything consequential, cross-venture, or outside scope). Authority is granted narrowly and revocably.

Inter-agent conflict resolution

When two agents disagree, the orchestrator arbitrates by evidence quality, confidence, and domain authority; an unresolved or high-stakes conflict escalates to human review. Agents never resolve consequential disagreements between themselves.

Error handling & recovery

Sandboxing & permissions

Agents run with least privilege. Legal, medical, financial, and security agents operate in restricted sandboxes with tighter data access and mandatory human approval. Permissions are deny-by-default.

Auditability & isolation

Every meaningful agent action is traceable to its inputs, permissions, and rationale. Cross-venture isolation is enforced: an agent built for one venture never leaks context or authority into another unless explicitly approved and logged.

Deactivation protocol

If an agent becomes unsafe, obsolete, or misaligned, any owner or governance authority can disable it immediately. Deactivation is instant, logged, and reversible only through re-registration.

Chapter EightAgent Quality & Learning

Quality dimensions

Beyond KPIs, every agent is evaluated on: accuracy, latency, reliability, citation quality, and human trust. An agent that regresses on any dimension is flagged for review.

The human feedback loop

Agents learn from corrections, not only outcomes. Every human edit, rejection, or override is captured as a training signal — the most valuable signal in the system (Law XIII).

Memory hygiene

Not every output becomes permanent. The card declares what is written to durable memory, what remains a temporary working artifact, and what is archived. Only validated, consequential lessons enter the genome (Book II); the rest expires.

Preventing regression

Every agent update is versioned and evaluated against a fixed benchmark set before release. A version that regresses is not shipped; a shipped version that degrades is rolled back to its last known-good state.