Why Industrial AI Agents Need Graph-First Orchestration

Note: This post has been rewritten under the May 2026 messaging. The original framing centered on per-device codegen for microcontrollers, which is no longer ForestHub's product story.

Five years ago, the open question for Edge AI was capability: could a model run on a small device at all? TinyML answered that. Today, the harder question is governance: can the agent that wraps that model be inspected, replayed, audited, and bounded — at the level the EU AI Act and the Cyber Resilience Act expect?

The Loop-First Default Doesn't Hold Up

Loop-first agent frameworks (LangChain, OpenAI Agents SDK, Claude Code) make the LLM's loop the program. For an assistant chasing open-ended goals, that's the right model. For a control system that decides whether a valve closes, it isn't. "Whatever the LLM decided" doesn't survive an audit, and no amount of sandboxes around the loop turns it into something that does.

Graph-First Orchestration

Industrial AI agents need a different architecture. ForestHub inverts the relationship: the graph is the program, the LLM is one node among many. Some nodes are deterministic (read a sensor, transform a value, branch on a condition, actuate). Other nodes are LLM agents. The engine orchestrates everything. The LLM is a component type, not the runtime itself.

What That Buys You

Inspectability — every possible LLM decision is a wire on the canvas. Replayability — record a production run, replay it in CI with the LLM mocked. Auditability — the action set is finite and enumerable. Boundedness — the LLM can only reach what's been explicitly wired in. Loop-first frameworks can't deliver these structurally.

Edge-Native by Architecture

The engine is a Go binary in a distroless Docker image (linux/amd64 + arm64) that runs on any Linux edge device — gateways, NUCs, NVIDIA Jetsons, Raspberry Pis, industrial PCs. Same image, same workflow, regardless of where it's deployed. Compliance (GDPR, AI Act, CRA) becomes a property of the architecture, not a policy bolted on after the fact.

Six concrete patterns this enables — verification-gated actuation, AI-as-classifier-only, replay testing, shadow-mode deployment, confidence-routed cascades, FSM-augmented agents — are documented at /patterns.

The graph is the program. The LLM is a node.