AI model, agent workflow and room control

In a controlled growing operation, three different jobs hide inside one phrase — "the AI controls the growing". Split them: deterministic control holds the climate inside agreed limits, an agent workflow works around the room with defined tools and permissions, and a model evaluates inputs inside that workflow. Each job is trusted on different evidence, and none of them borrows another's authority.

Deterministic control holds the room

A growing room's climate is held by a closed control loop: conditions are sensed, a control decision is made, and HVAC, heating and ventilation equipment actuates — continuously. Greenhouse-climate research treats this as one integrated loop, from sensing and data governance through control strategies to deployment (Wang et al. 2026, MDPI Agriculture). Rule-based control remains a standard layer in that loop, including safety fallback, even when model-based methods are added on top.

Trust here rests on testing, not intelligence: control logic tested against the agreed limits, defined behaviour when a sensor or actuator fails, and an audit trail of setpoint changes. This job asks no permission from any model.

An agent workflow works around the room

An agent can organise work around the room without operating it: read permitted records for a stage, compare a room against its schedule, assemble a report, prepare a draft for a person. Anthropic's engineering guide draws the distinction this article uses: a workflow is a system where LLMs and tools are orchestrated through predefined code paths, while an agent directs its own process and tool use (Anthropic, Building effective agents).

An illustrative reporting task: retrieve one room's climate history for one week and the operator's notes for the same period, then prepare a summary draft. Preparing it is not sending it, and describing a deviation is not changing anything.

The evidence for trust is the boundary itself: a defined tool list, defined permissions, a tested scope and human acceptance before production. Outputs that carry unrestricted authority over a production system have not earned trust; they skipped the test.

A model evaluates inputs

Inside that workflow, a model classifies, estimates or drafts text from sensor and process data, summaries and reports. Its output needs an interpretation and a defined consumer. Producing a recommendation is not authority to actuate equipment. In safety-critical contexts, LLM reliability, accuracy and trustworthiness remain critical concerns (Dokas et al. 2026, Safety Science); evaluation and human oversight are needed before outputs are trusted (Anthropic, Building effective agents; Wang et al. 2026).

Keep three arrows separate

On one diagram, mark three arrows differently: information, recommendation, authorised action. Then ask what evidence and permission let each arrow exist, and who reads the result:

Which function holds the climate limits, and where are its tested logic and audit trail?
Which function prepares reports, with which tools and permissions?
Which function may change a setpoint, and who reviews a recommendation before it does?

If the three answers come from three different parts of the architecture — and a person sits between recommendation and setpoint change — the description is specific enough to check. How to allocate such responsibilities between edge and central services is a related, separate decision (edge and cloud IoT responsibilities).

Manna MIND is built around this separation. Manna configures the MIND Agentic Server around an existing growing system, or builds a new system on the full-stack MIND AIoT platform: agents work with defined permissions and are tested and accepted before production, while deterministic control keeps holding the room. In Manna's usage, AIoT means Agentic IoT. See the Platform page.

Prepared with AI assistance for Manna Insect Oy.