An agent task brief that defines the finish

Before an agent starts work in a growing operation, write the finish into its brief. A useful brief defines a result that can be accepted or refused: one named result, the records and window the agent may use, the actions it may take and must refuse, the evidence of completion, and who accepts. "Look at the climate data and improve things" names none of these — the guesses end up in a summary nobody can check, or a change nobody authorised.

Vendor guidance points the same way: agents need clear success criteria, human checkpoints and explicit stopping conditions (Anthropic), and instructions should cover what to do when required information is missing (OpenAI).

One named result — preparation, not application

Name the result as a checkable object: a document, a list, a draft decision. Say whether the agent prepares it or also applies it — sending a report and changing a setpoint carry different authority. Tools define functional boundaries, so choosing what the agent cannot do is part of the design (Anthropic).

Permitted records and window

List the sources, rooms, batches, stages, period, timezone and units the agent may use, and say what is out of scope. When a source is missing or incomplete, mark the gap rather than smooth it into a narrative.

Allowed and refused actions

Reading history, writing a draft file and changing a live setting are different kinds of action. State which are allowed, which need approval and which are refused. Where the consequence is high — equipment, environment, outgoing messages — do not rely on the model behaving well: limit access so the refused action cannot be taken.

The finish and stopping rules

Define completion so a reviewer can check it without re-reading the agent's prose. For longer work, agree stopping rules — a maximum number of iterations, or a checkpoint reporting coverage, findings and gaps before continuing (Anthropic). A justified partial or blocked result is a real outcome: "I could not finish because the logger was offline" arrives as a result, not as silence.

Who accepts, and on what basis

Name who accepts — in a growing operation, the technical operating owner of the installation — and the basis: each main conclusion traces back to the permitted records, independently of how fluently the result is written. A fluent paragraph is a claim, not proof; the acceptance check is the habit described in our article on specifying grow-room reports.

An illustrative brief

As an explicitly illustrative example, not a real case or measured result: a weekly climate-deviation summary for one room and one stage.

Result: a summary listing deviations between recorded climate values and the grow schedule for that room and stage, each traced to its log entries; the agent prepares it and changes nothing.
Records: the climate logs and grow schedule for that room, stage and week only; a missing log entry is shown as a gap.
Actions: reading and summarising are allowed; changing a setpoint, adjusting equipment and sending the summary are refused.
Finish: every day of the window is either summarised or explicitly marked missing, and the report says which.
Acceptance: the grower accepts after tracing each listed deviation to the records.

The example stays method-level: no species-specific setpoint, deviation limit or threshold — those belong to your own process. Batch and stage context turns sensor numbers into an operational answer (our article on process context), and agent work stays separate from deterministic control (our article on AI models, agents and workflows).

The brief, in one list

one named result, with preparation separated from application
the permitted records, window and missing-source rules
allowed and refused actions, enforced where the consequence requires it
the acceptance evidence and stopping rules, with a justified partial or blocked result acceptable
who accepts the result, and on what basis

Write it short enough that a human operator could execute it too.

If you would rather have this configured for you: the Manna MIND Agentic Server is configured around your growing system for Manna-run agent analysis and reports, and you stay the person who accepts the result. To talk through your first task brief and the records it may use, contact Manna.

Prepared with AI assistance for Manna Insect Oy.