We don’t hide the architecture.
We sell the architecture.
Most agencies treat their methodology as the moat. We treat it as the marketing. The agent stack, the rubrics, and the architectural prompts are published quarterly. We compete on execution quality and the depth of our memory of your business — not on a black box.
Twelve agents.
Each one is five things.
Each agent in the system is not a chatbot. Each agent is the same five-component object:
- System prompt — explicit role, mission, output requirements, boundaries, voice canon.
- Curated tool kit — the APIs, scrapers, databases, and document stores the agent is allowed to call.
- Memory store — scoped to the agent’s role and the active client account; isolated across accounts.
- Evaluation rubric — specific to the agent’s outputs; produces a score and a verdict on every artifact.
- Deterministic handoff protocol — the structured contract by which the agent passes work to other agents.
Agents do not freelance. They do not improvise structure. Their inputs and outputs are JSON-typed against published schemas. When an agent’s confidence is below threshold, it produces an escalation ticket instead of a deliverable.
Sentinel runs.
On every output.
Rubric pass
Every artifact — a draft page, a sequence, a research pack, an outbound message — gets scored against an output-specific rubric before it leaves the system. Items below threshold are routed back to the producing agent with specific edit notes, or escalated to the operator.
Source-of-truth check
Stricture (the editor) verifies every factual claim against an approved source set: the client’s product docs, public competitor pages, primary research, the case-study source map. Unsourced claims do not ship.
Confidence collapse
Sentinel monitors confidence over the trajectory of a workflow. When two consecutive agents flag uncertainty, the workflow pauses and an escalation ticket is opened. We do not ship under uncertainty; we ask.
Tone and IP
Sentinel enforces voice adherence and watches for IP risk — uncited near-duplicates of competitor copy, regulated-advice patterns, defamation surface area. Catches that fail this gate are tracked and published in the quarterly Field Note: “What Sentinel caught this quarter.”
How work moves.
Through the system.
Atlas → Cartograph → Lyric → Stricture → Tessera → Forge → Ledger → Concord
Quality bar: unique value, source-backed claims, correct CTA, schema/canonical/internal links/indexation plan. Pauses when source quality is weak, regulated advice surfaces, claims are unsubstantiated, or similarity to competitor content is too high.
Ledger → Vellum → Lyric → Stricture → Sentinel → Concord
Use cases: onboarding, activation, win-back, expansion, transactional copy. Acceptance gates: explicit segment and trigger, one job per email, no fake urgency, approved claims, metric defined before launch.
Atlas → Cartograph → Beacon → Stricture → Sentinel → Concord
Rules: no spam-cannon behavior; every message has a real signal; Beacon drafts only; the operator and client approve send policy; warm replies route to Concord then to the client’s sales owner; suppression lists are honored.
Ledger → Concord → Operator → Client
URL-level attribution, channel-level pipeline, anomaly notes, decision requests. Drafted automatically, refined by the operator, shipped on Friday. Public dashboard refreshed daily.
What agents will not do.
Without explicit operator approval.
- Publish to a client’s CMS without operator review on the first run.
- Send outbound messages without an approved send-policy SOW.
- Sign, invoice, refund, or move money. Quill drafts; the operator approves.
- Delete client data, revoke access, or mutate production systems beyond the agreed scope.
- Generate content for regulated-advice domains without explicit client sign-off and source provenance.
- Override Sentinel. Sentinel can pause any workflow.
Things we publish.
On purpose.
Quarterly methodology Field Note
Architectural-level prompts, rubrics, and workflow updates from the previous quarter. We compete on execution and memory, not secrecy.
Public attribution dashboards
Atelier and Bureau clients receive a daily-refreshed dashboard with URL-level pipeline, channel attribution, and anomalies. We can also publish anonymized snapshots when clients consent.
escalation and correction protocol
A documented escalation and correction protocol. If something goes out wrong, ownership, notification, revision, and prevention steps are already written down.
Subprocessor list on request
Vercel, Anthropic, Google Workspace, Stripe, and the client’s CRM. Updated when material changes occur. Provided to any prospect or client who asks.
No single-vendor dependency.
Models picked for the job.
AXIOM runs on a hybrid model stack: premium AI subscriptions where speed and judgment matter, local models where privacy or cost margin matter, and deterministic automations where reliability matters. The architecture is explicit about which work runs where.
Premium frontier (subscription). frontier models for strategy, adversarial editing, final QA, and complex customer-facing judgment. The exact provider can change as models improve; the architecture is designed so the workflow survives vendor changes.
Local models (private, low-marginal-cost). approved local or private models for low-risk drafts, internal preprocessing, and customer work where data sensitivity, resilience, or cost requires it. Local models never run client-facing, legal, payment, or security decisions on their own — those always route through frontier judgment with Sentinel oversight.
Deterministic automation. n8n / Temporal / Windmill orchestration for handoffs, retry, audit logging, and circuit breakers. Python and SQL where math and data joins are involved. Anything that can be deterministic, is — LLMs are reserved for judgment, not arithmetic.
The result: lower marginal cost than agencies that pay 12 salaries, lower vendor risk than firms locked into a single provider, and architectural transparency that lets a sophisticated buyer evaluate the engine, not just the demo.