Loop + Graph EngineeringThe control surface for living systems of work

Discipline

What is Loop Engineering?

Loop Engineering is the practice of designing systems that prompt, verify, and stop AI agents for you, instead of prompting them turn by turn.

You no longer sit in the inner loop. You design the outer system that:

• Discovers work
• Hands it to an agent (or set of agents)
• Checks the result with independent verification
• Persists state outside the conversation
• Decides whether to continue, escalate, or stop

The core shift was named and popularized in mid-2026 by Peter Steinberger, Addy Osmani, and Boris Cherny (Claude Code). Cherny’s practice is representative: he no longer prompts the model directly; he writes loops that do.

Five building blocks + memory

Discipline

What is Graph Engineering?

Graph Engineering is the design of agents, deterministic code, tools, humans, and evaluators as nodes, together with execution conditions, dependencies, state transitions, parallelism, retries, termination, and permissions expressed through edges and shared state.

It is the control surface that connects inputs, processing units, and review.

Graphs contain loops. They do not replace them.

A single agent still runs an inner loop. Multiple specialized agents or steps are wired into an explicit topology so that state, permissions, and failure paths are managed at the system level rather than inside one conversation window.

Metabolic Organizational Design

Why both matter for Metabolic Organizational Design

Traditional organizations were designed as linear value chains. Metabolic design treats the organization as a living system that must convert information and action into coherent adaptation under entropy.

Loops give individual work units the ability to iterate, verify, and improve without constant human prompting. Graphs give the larger system topology, shared state, parallel paths, and explicit hand-offs between Context Creators, agents, and judgment infrastructure.

Together they support:

• The Information–Action Cycle at multiple scales
• Permanent scaffolding for judgment (Strategy Engineer and related agents)
• Fractionalized, network-first work across independent Context Creators and Builders
• Systems of Cognition that remain coherent as agentic AI is adopted at scale

The substrate

SurrealDB for compounding memory and knowledge graphs

Agents fail more often from context and memory problems than from model limitations.

A durable substrate must support:

• Structured entities and rich, typed relationships (true knowledge graphs)
• Vector embeddings co-located with the data
• Hybrid queries (graph traversal + semantic search + structured filters) in one engine
• Real-time updates and ACID consistency so multiple agents can share state safely
• Temporal and multi-hop reasoning without stitching five different stores together

SurrealDB is purpose-built for this. It is a multi-model database (graph + document + vector + relational + time-series + more) with a single query language (SurrealQL) that makes knowledge-graph patterns native rather than bolted on.

We use and recommend it as the context and memory layer beneath loops and graphs.

Featured SurrealDB resources

Getting started inside the network

How to begin

1. Begin with a single bounded loop that has explicit halt conditions, independent verification, and external memory.
2. Only introduce graph topology when you need multiple specialized nodes, parallel paths, or shared state across agents.
3. Treat state schemas, checkpoints, and permissions as first-class design work — not afterthoughts.
4. Keep human authority at high-stakes decision points. The graph and loop should surface evidence; judgment remains human where it matters.
5. Prefer a durable multi-model store (SurrealDB) over fragmented vector + document + graph stacks so context compounds instead of drifting.

The Value Network

Metabolic Organizational Design for the Intelligence Age