What agents mean for learning design
Microsoft Copilot Studio agents are AI-powered tools that carry out multi-step tasks autonomously — triggering workflows, querying data sources, generating documents, and taking actions in connected systems. They are not chatbots. They act.
This changes what "using AI" means in the workplace. With Copilot, the human reads and evaluates an output. With an agent, the agent may have already sent the email, updated the record, or triggered the approval before the human sees the result. The stakes of a misunderstood brief are higher. The learning design challenge is correspondingly harder.
By the end of this unit- Describe the difference between a Copilot interaction and an agent action, and its implications for learner accountability.
- Identify the three prerequisite capabilities a learner needs before working with an agent.
- Design an agent onboarding module structure that addresses trust calibration and error correction.
The design shift from Copilot to agents
- The human reviews the output before anything happens externally
- The output is text — a draft, a summary, a list
- Errors are corrected before use
- The human's judgment is the final gate
- The agent may act before the human reviews — sending, updating, triggering
- The output is an action in a connected system
- Errors may need to be reversed in a downstream system
- The agent acts within the permissions it has been granted
- Learners need to understand what the agent is authorised to do before they trigger it
- Trust calibration is a prerequisite, not an advanced topic
- Error correction needs to be designed into the onboarding — not left to incident response
- The "who is accountable" question becomes more urgent when the AI acts
Three things learners need first
Before a learner is ready to work with an agent in a live system, they need three prerequisite capabilities. These should be confirmed — not assumed — before the agent onboarding begins.
1. Understanding of what the agent is authorised to do
The learner must know which systems the agent can access, what actions it can take, and — critically — what it cannot be asked to do. Without this, the learner cannot assess whether the agent's proposed action is appropriate, or identify when the agent is about to do something unintended.
Assessment approach: Present a list of tasks and ask the learner to classify each one as within or outside the agent's authorisation. Include edge cases.
2. The ability to write a brief that the agent can act on accurately
An agent that receives an ambiguous instruction will either ask a clarifying question (good), make an assumption (risky), or produce an error (recoverable, but disruptive). The learner must be able to write briefs precise enough to avoid the second and third outcomes.
Assessment approach: Give the learner a task scenario and ask them to write the instruction they would give the agent. Evaluate for completeness, specificity, and the absence of ambiguous terms.
3. Knowledge of how to stop, review, and correct an agent action
Before going live, the learner must know: how to pause an agent mid-task, how to review what the agent has done in a session, and how to initiate a correction in the downstream system if an error occurred. This is not advanced material — it is day-one safety knowledge.
Assessment approach: Walk through a scenario in which the agent takes an unintended action. Ask the learner to identify the steps to correct it, in the right order.
These three capabilities are not aspirational stretch targets — they are go/no-go criteria for working with a live agent. Design your pre-agent assessment to confirm all three before the learner accesses any agent in a production system.
Designing agent onboarding
Agent onboarding has a different structure from general Copilot onboarding. It must move the learner from awareness to safe, independent operation — not just feature familiarity. The recommended structure is a four-stage progression that matches the trust-building requirement.
Stage 1 — What this agent does (and what it does not)
A precise description of the agent's capabilities, authorisation scope, and the systems it connects to. Include explicit examples of tasks that are within scope and tasks that are not. Learners who do not know the boundary cannot respect it.
Stage 2 — Watched demonstration
A screen-capture demonstration of the agent completing a standard task, including what the agent shows the user before acting and what the confirmation step looks like. Narrate the decision points, not just the clicks. Show one near-miss: a scenario where the agent was about to take an unintended action and the user caught it before confirming.
Stage 3 — Supervised practice in a safe environment
A sandbox or demo environment where the learner can trigger the agent without affecting live systems. Design at least two tasks: one where the agent performs correctly and the learner confirms; one where the agent makes a plausible error and the learner must catch it before confirming. Both require the accountability check before submission.
Stage 4 — Error recovery exercise
A scenario in which an error has already occurred in a downstream system. The learner must identify what happened, find the correct reversal action, and complete it. This stage is non-optional. Error recovery must be practiced before it is needed.
Think of a specific agent your organisation is deploying or considering. Can you complete Stage 1 of the onboarding design above — listing precisely what the agent can and cannot do? If you cannot, that is the first thing the design needs.
Trust, verification, and error correction
Trust calibration is one of the most underdesigned elements of AI onboarding — across both Copilot and agents. Both over-trust (accepting all AI output without review) and under-trust (refusing to use the AI because of fear of error) are failure modes. Learning design should aim for calibrated trust: using the AI confidently within its reliable range and verifying at the boundaries.
Designing for calibrated trust
Show the reliability range, not just the capability
Build verification into the standard workflow
Make error correction destigmatised and practised
Anchor accountability — always
A learner completes the agent onboarding and says "I trust the agent completely — it never makes mistakes in the demo." What is the instructional design problem here?
End of Unit 6
You should now be able to:
- Distinguish the design requirements for Copilot onboarding versus agent onboarding.
- Identify the three prerequisite capabilities for safe agent use and design assessments for each.
- Apply the four-stage agent onboarding structure, including error recovery practice.
Unit review
What is the most significant design difference between Copilot onboarding and agent onboarding?
Which of the three agent prerequisites is most commonly missing from agent onboarding programmes?
Why must error recovery be practised in a safe environment before the learner goes live?
What does calibrated trust mean in the context of agent use?
End of module
You have completed Course 06: Building Agent-Ready Learners. Next: Briefing AI for Learning Media — using AI generation tools to accelerate audio, image, and scenario production in eLearning.