Build, Rent, Skip: The L0–L8 Decision Matrix Every Agent Team Eventually Writes.

Every agent team I talk to is having some version of this debate, often all four at once, often without realizing they're four different layer-level decisions wearing the same coat. The fastest way out of the debate is to stop arguing about the components and start naming the layer each component lives on. Then apply a default.

Supply Chain of Intelligence™ (SCoI) gives you the matrix. The defaults below are the answers most teams should give most of the time, with the explicit conditions for when to deviate. Same canonical definition as always: intelligence is a supply chain, value accrues at the bottlenecks. That sentence implies most of the matrix on its own.

The default matrix

Ten layers, three columns, one default per cell. Read this once and put it on the wall.

L-1 Resources → Rent. You are not running a fab or signing a 400MW PPA. The exception is hyperscalers and the very largest model labs, and if you're reading this, that's not you.

L0 Infrastructure → Rent. Cloud, GPUs, edge. The build-vs-rent debate here is a distraction in 95% of cases. Even fast-growing model companies are renting most of their compute through L+ year three.

L1 Data → BUILD (the proprietary slice). This is the most important cell in the matrix. L1a (public/open) is rented. L1b (proprietary), L1c (behavioral), L1d (outcome) are built, always. If you don't own at least one of L1b/c/d, the framework's prediction is that the platform layer below you absorbs your product. Law I.

L2 Models → Rent base, build fine-tunes only after L1 is clean. Foundation models are rented. L2b fine-tunes are built only when L1b gives you a corpus other people can't reproduce. L2c embeddings are rented unless your domain has weird tokenization. L2d routing is not a build, it's a runtime concern that lives in L6.

L3 Gatekeeping → GATE (build verification, rent compliance tooling). L3b quality gates and L3c provenance are almost always built in-house, because they encode your business's risk tolerance. L3a compliance tooling (SOC2 evidence, audit logs, eval harnesses) is rented from Vanta-style vendors. Law IV is non-negotiable in regulated spaces.

L4 Access → Rent the pipes, build a thin facade. Every vendor API (Stripe, Salesforce, Slack, MCP servers) is rented. The facade in front of them is built, because the facade is what protects you from vendor churn and what makes surface-swapping a local change.

L5 Execution → BUILD. This is your actual product. If L5 is rented (i.e., you're a thin wrapper over someone else's skill), you don't have a company; you have a feature. L5a tool use, L5b reasoning scaffolds, L5d operating playbooks, all built. L5c RAG plumbing can be partially rented.

L6 Orchestration → Rent the framework, build the policy. Use LangGraph, CrewAI, or your own thin wrapper for the loop primitives. Build the policy that lives inside it, role routing logic, context selection, retry/escalation rules. The framework is infrastructure; the policy is IP.

L7 Surface → Ride existing habits before building new ones. Slack, email, Teams, Notion, the user's IDE, the user's browser, the user's CRM. Build a new surface only when no existing surface fits the modality (voice agents, ambient devices). Law III warning applies: surface without depth is a graveyard.

L8 Memory → BUILD, start thin in v1. Session memory and entity profiles in v1, schemas for institutional knowledge designed in v1. L8 is the most under-built layer in early agent products and the layer with the highest compounding return. Don't skip it because it doesn't show in the demo.

The three places teams get it wrong

Out of dozens of agent teams I've talked to in the last year, the same three mistakes show up regardless of vertical or team size. Each one is a misread of which column a layer belongs in.

Mistake 1: Building L0 or L2 to “control the stack.” The instinct is understandable. Cost optimization, latency, perceived lock-in. The framework reads it as a waste of engineering hours. Law II says value accrues at bottlenecks, and for an agent product the bottlenecks are L1b and L8, not L0 or L2. Every engineer-week spent on L0 infra is a week not spent on the layers that actually compound. The exception is companies whose business is L0 or L2 (NVIDIA, Anthropic, OpenAI), which is again, not you.

Mistake 2: Skipping L3 in v1. “We'll add compliance and verification when we have customers.” The framework predicts you won't get the customers without it. Any agent that touches money, regulated data, or customer-facing communication needs L3b and L3c at launch. The work is small in v1 (logging, a verification step on writes, provenance traces). The work is enormous if you bolt it on after you've shipped. Law IV doesn't take a quarter off.

Mistake 3: Treating L8 as a v2 problem. Memory is the only layer where the cost of waiting is asymmetric. The schema decisions you make in v1 determine whether L8d (institutional knowledge) is buildable in v2 at all, or whether you have to do an expensive data migration first. The smallest viable L8 in v1 is a database with a clear entity model and event log. That's a week. Skipping it costs a quarter later.

The matrix is the conversation

The reason this matrix is useful is not that the defaults are surprising. They aren't. Most senior builders would arrive at most of these cells on their own. The matrix is useful because it gives the team one vocabulary to argue inside of. Instead of debating components (Pinecone vs. pgvector, LangChain vs. CrewAI, OpenAI vs. Anthropic), the team debates layers (do we own L2c, do we build the L6 policy, what's our L3 gate). The component choice falls out of the layer decision, almost always.

Pin the matrix to the wall. Mark your current state in each cell. Mark your six-month state. The delta is your roadmap. The framework didn't invent the answers, it organized them so the team could stop re-debating the same question every sprint.

Same definition every time: intelligence is a supply chain. Build at the bottlenecks. Rent the rest. Gate where the law of separation demands it. Skip nothing at L8.

The full framework, ten layers, fifty sublayers, four laws, and the Agent Decoder, is free at supplychainofai.com. No signup, no paywall.

- Anand