Lesson

G3 Write an AI Agent Task That Can Be Reviewed

On this page

Turn Intent into a Defined Task

“Optimize my schedule” omits data, permitted actions and evaluation criteria. A reviewable Agent task identifies the sample, allowed changes, required outputs and human confirmation points.

Use the standard injection sample in two stages. First prepare a configuration diff that changes only the color weight, together with pre-run checks. After approval, an authorized runtime executes the comparison. Save configuration proposals separately from run results.

Example Task Definition

Item	Content
Task	Compare one color-weight change with the same data and constraints
Inputs	Confirmed production_scheduling version and field definitions
May change	The color weight, with its value specified by the requester
Keep fixed	capacity weight, work center, 40×3 structure, constraints and data
Resource limits	Read from the runtime being used
Stage 1 output	Configuration diff and pre-run checks, clearly marked as not yet run
Stage 2 output	Actual run status, result comparison and interpretation
Human review	Before execution and before operational adoption

Keep authorization files in the execution environment; the task package only needs information for preparing the configuration. Confirm what data may be shared before using an external Agent service.

Ask Specific Questions

Have the Agent explain which business attribute color maps to, what capacity means in this recipe, which records prove execution, how to compare different weights, and how to describe the absence of a usable plan.

These questions encourage use of the structured inputs rather than assumptions from general scheduling knowledge. If capacity is interpreted as machine availability, correct the interpretation and configuration before continuing.

Distinguish Three Outputs

Type	Label and review
Configuration proposal	Not executed; inspect parameters and differences
Execution request	Record recipient, input version and computing limits
Run result	Actual status, actions and result evidence

Match the request to its run identifier, then inspect candidates, the selected solution and the source of each measure. Trace the result explanation to the actual configuration and output records.

Where a workbench supports task-package export, pass the task, context, expected results and permitted actions to the chosen Agent. A user or configured integration then invokes it and reviews the response. Export prepares the information; invocation is a separate step.

A Task Prompt to Adapt

This is a natural-language task for an Agent, not an executable API request. The requester supplies the bracketed values. Resolve missing values before proceeding.

Task: Prepare an injection-scheduling comparison that changes only the
color objective weight. Do not start solving yet.
Inputs: data version [supply]; production_scheduling recipe version [supply];
field definitions [provide]; active runtime parameter limits [provide].
Change: color from [baseline value] to [requested value].
Keep fixed: data, constraints, capacity weight, work center, sequence length,
cycle count, thread count and time limit.
Check: in this sample, color maps to property_1; the capacity objective
relates to container volume in property_3, not machine capacity.
Return: baseline/proposed configuration diff, parameter checks, missing
information and questions requiring confirmation.
Pause if a field is unknown, a limit is exceeded or a fixed item must change.
After approval, the prepared runtime runs baseline and adjusted configurations.
Only then report actual status, sequences, both change counts and objective measures.

For a reading check, reject a proposal that also changes the capacity weight. Flag a missing requested color value for confirmation. If all checks pass, the configuration can go to the person responsible for execution. None of those statuses establishes an improved result before the runs occur.

Exercise and Review

Write a task and two rejection conditions: a request outside permitted parameter limits, and an attempt to change unauthorized data or fields. Add a normal acceptance condition: only the specified weight changes and all fixed items remain consistent.

Submit the task, configuration diff, checks and status explanation. The Agent helps organize work; enterprise users use the evidence to decide whether the configuration and plan can be adopted.