A pattern library for designing AI coaches, tutors, copilots, and reflective agents for workplace learning.
This repo collects reusable design patterns for agentic learning systems — learning experiences where an AI agent plays an active role: diagnosing, coaching, simulating, questioning, or measuring.
Each pattern is a self-contained brief you can hand to a designer, an engineer, or an LLM. It names the learner need, the agent's role, the interaction loop, the data required, the risks, and the signals that tell you whether it's working.
Who this is for: learning designers, people scientists, L&D strategists, and engineers building AI-enabled learning inside organizations.
An agentic learning system is not a chatbot bolted onto a course. It is a system where the AI holds a defined role in a learning loop — it observes something about the learner, acts on that observation, and adapts based on the result. The unit of design is the loop, not the conversation.
The distinction matters because most "AI in learning" today is content generation: faster courses, more quizzes. These patterns aim at something different — behavior change in the flow of real work.
| # | Pattern | Core learner need |
|---|---|---|
| 01 | Diagnostic Coach | "Where am I actually weak?" |
| 02 | Socratic Tutor | "Help me think, don't tell me the answer." |
| 03 | Practice Simulator | "Let me rehearse before it counts." |
| 04 | Feedback Analyst | "Make sense of the feedback I'm getting." |
| 05 | Manager Coaching Copilot | "Help me have better coaching conversations." |
| 06 | Reflection Journal Agent | "Help me notice what I'm learning." |
| 07 | Skill Transfer Agent | "Help me use this in my actual work." |
| 08 | Learning Measurement Agent | "Tell us whether any of this changed behavior." |
- Start from the learner need, not the technology. Pick the pattern whose "core learner need" matches your problem.
- Adapt the interaction loop to your context — the loops are deliberately concrete so you can see what to change.
- Check the data requirements early. Most agentic learning projects fail on data access, not model capability.
- Steal the evaluation signals. Every pattern ships with signals that indicate behavior change, not just engagement.
- Use the example prompt as a starting point for prototyping in any capable LLM, then harden it.
To propose a new pattern, copy templates/pattern-template.md and open a pull request.
Completion data tells you something happened. It cannot tell you whether capability changed, whether behavior transferred to real work, or whether the work improved. Every pattern in this library therefore includes evaluation signals — observable evidence, in the flow of work, that the loop is producing change. If you can't name the signal, you're not ready to build the agent.
- Add worked examples (transcripts) for the three most-used patterns
- Add an anti-patterns file: common agentic learning failures
- Add evaluation rubrics that pair with learning-measurement-playbook
MIT — use these patterns anywhere. Attribution appreciated, not required.
Chris Richardson — chrisrichardson.dev · LinkedIn