Nodes
Agents, deterministic guards, human decisions and durable artifacts each perform a bounded role.
A naïve-labs working definition · 2026
Graph Engineering
Design the paths agents can take — and the evidence, state and authority that travel with them.
working graph live
learning edge open
One working graphFlaiwheel / 2026
sourceMemexSource identity, readable meaning and provenance.
semantic nodeLotseSummarizes and assesses bounded project relevance.
typed handoffNutrientCompact evidence crosses the boundary without editing local truth.
scope-owning nodeGardenerClassifies and integrates inside the project boundary.
Flower / Petal
Seed
Activity
Resource
Governance proposal
Compost
01 / Definition
The graph is the operating structure.
Loop engineering shapes how one agent observes, acts, checks and repeats. Graph engineering shapes how loops, deterministic steps, human gates and durable artifacts connect into a governed system.
Edges
Every transition states what may move, what must be checked and where work can stop.
State
Handoffs, receipts and provenance survive individual model calls and make the graph inspectable.
Authority
The topology carries permissions: one node may assess, another may write, and high-gravity paths remain human-gated.
02 / Working graph
From captured meaning to local substance.
Flaiwheel already runs a narrow graph across two semantic agents and several deterministic boundaries.
Lotse lands cargo
Lotse reads bounded source meaning, scores project relevance and can select one direct target only under a conservative evidence gate.
A Nutrient crosses
The handoff carries compact facts, questions and provenance. It is deliberately not a task, decision or project update.
Gardener integrates
Gardener reads the local project context, chooses an enabled route and records the resulting transition with a receipt.
03 / The handoff
A boundary object, not a message.
The Nutrient is the graph's most important edge made tangible. It lets useful meaning move between agents without granting the upstream agent authority over the downstream domain.
Lotse may decide that material belongs at a project dock. It may not decide whether that material becomes a Seed, Flower, Activity, Resource or governance proposal. That interpretation belongs to Gardener inside the project's native scope.
Deterministic guards verify project maps, evidence identities, routes, gravity and idempotence around both semantic judgments. The graph therefore carries more than data: it carries epistemic and operational limits.
04 / Distinctions
Loops live inside graphs.
Loop engineering
Designs one repeated cycle: observe, decide, act, verify, retry or stop.
Graph engineering
Designs how multiple loops, roles, states, branches and authority boundaries compose.
Knowledge graph engineering
Models entities, facts and relationships. It may support an agent graph, but it is not the same problem.
05 / Current status
The graph works. The learning edge is open.
One real path has crossed from Paperless through Memex and Lotse into a project Nutrient, then through Gardener into a provenance-preserving Resource.
Gardener's outcomes do not yet recalibrate Lotse automatically.
Relevance assessments and route receipts already create evidence for learning. The feedback path from those outcomes back into Lotse's thresholds is not yet closed. Until reviewed examples are sufficient, autonomous landing remains deliberately restricted to one exclusive direct fit with strong evidence.
This is not hidden implementation debt. It is the next research question: when does downstream success become trustworthy upstream calibration without erasing local judgment?
06 / Practice
Graph engineering makes allowed movement as explicit as capability.
A diagram can show that agents are connected. An engineered graph explains why a transition is allowed, what evidence supports it, where state persists, who may reject it and how the system learns without pretending to know more than it does.