How the principle is rendered into systems — and the rooms in which the work happens.
Sovereign-recursive systems are constructed in five layers, each with its own invariants. Mutation flows from the top; the lower layers are sovereign over the upper. L4 holds the gate. L5 holds the eval. Neither is reachable from L3.
┌──────────────────────────────────────────────────────────┐ │ L5 — Continual evaluation │ │ Capability evals, operator evals, regression suite │ ├──────────────────────────────────────────────────────────┤ │ L4 — Gating, rollback, jurisdictional firewall │ │ Apply-then-evaluate. Mutation has no write access │ │ to L1, L2, or L5. │ ├──────────────────────────────────────────────────────────┤ │ L3 — Mutation operators │ │ Prompt, tool description, sub-agent topology │ ├──────────────────────────────────────────────────────────┤ │ L2 — Self-observation and trace │ │ Append-only. Immune to L3. │ ├──────────────────────────────────────────────────────────┤ │ L1 — Kernel control loop │ │ Async tick. Framework-agnostic. Sovereign. │ └──────────────────────────────────────────────────────────┘
The full anchor essay walks through each layer with working code: Sovereign Recursion: Why Your Self-Improving Agent Needs a Jurisdictional Firewall →
Four ways to bring this architecture into an organization. Each tier is fixed-scope, fixed-price, and operator-led. Engagements are direct with Justin Oberg as principal — not handed to junior staff after sale.
A deep, structured review of your existing agent stack against the L1–L5 model and the jurisdictional-firewall principle. The deliverable is a written report — capability gaps, layer violations, the three highest-leverage changes you could make this quarter, and an honest assessment of whether further engagement is the right next step.
The lead surface for engagements. Most teams who later commission a build start here.
End-to-end delivery of one specific architectural piece. Common shapes:
Fixed-scope, fixed-price, fully remote with one optional in-person workshop. Working code shipped at the end.
Bespoke design and delivery of a complete L1–L5 sovereign-recursive system for a team deploying agents in production. The deliverable is the running stack — not a deck. Includes:
The codebase and the data are yours. The system is maintainable by your engineers after delivery. Three to four of these per year.
Ongoing architectural advisory for engineering teams already running recursive agents. Async by default. Includes:
For teams who have shipped past the build phase and need ongoing judgment, not a fresh build. No on-call. No SLA. Six seats, total.
A 30-minute conversation, free, by appointment. Two purposes: confirm the architecture is the right answer to your problem, and confirm we can actually work together. Some firms walk away knowing they don't need this yet — that is a fine outcome.
If we proceed: a one-page written proposal with deliverables, timeline, and price. Fixed-scope. Fixed-price. No hourly billing for build work.
Remote by default with one or two in-person workshops at your offices. Weekly written progress updates. Working artifacts shared as they ship.
The deliverable is a working system, not a deck. Documentation and methodology transfer to your internal team. The codebase and the data are yours.
If the architecture is the right answer to a problem you're carrying, the conversation is worth having.
No deck. No follow-up automation. No pitch.
Engineering teams shipping recursive AI agents in production at growing organizations — defense-aligned, federally-funded research, mid-stage software companies, and mid-size firms whose agent deployments have reached a scale where ad-hoc gating is no longer sufficient. I do not take engagements with companies whose AI deployment is primarily content-generation, marketing automation, or consumer chatbot work.
Yes to NDAs. Yes to regulated environments (HIPAA-aware, financial, controlled-unclassified information). Classified work specifically requires a separate conversation about clearance scope and may route through a different vehicle. Federal-aerospace background; familiar territory.
Architecture reviews: typically within two weeks of proposal acceptance. Focused engagements: 2–6 weeks lead time. Full engagements: 4–10 weeks lead time, depending on calendar. Three to four full engagements per year is the realistic ceiling.
Then we don't, and the conversation ends amicably. The diagnostic exists specifically to surface that — misfit engagements are bad for both parties. If a different practice would be a better fit for your situation, I will say so.
Possibly later. The current focus is principal-led engagements. The essays remain free and continue to publish on a slow cadence.