Set the limits of autonomy
Relate what the agent can do to the competence students must build, the risk of the task and the human review actually available.
Fernando Santamaría · University handbook · 2026
When AI begins to act,
what should the learner still do?
A book about understanding, designing and governing AI agents without losing sight of what it means to learn.
Agentic AI in Higher Education
Foundations, architectures, pedagogical design and governance of agents in higher education.
20chapters
7annexes
24pages in the toolkit
2languages available
The question running through the book
Delegate operations.
Preserve judgement.
A system’s operational autonomy and a learner’s agency are not the same thing. A tool’s ability to complete a task is not enough to decide whether it should do so in a learning activity.
The book connects technical architecture with concrete decisions: what support to offer, what effort to preserve, when to intervene and how to check whether students can explain and transfer what they have learned.
For university educators, learning designers, researchers and teams deciding how to bring AI into their institution.
Decisions the book helps you make
Three decisions that no longer need to be guesswork.
Evaluate before adopting
Separate technical feasibility from educational impact, place applications according to the available evidence and document why they are accepted or rejected.
Check what was learned
Look beyond the submitted product through evidence of process, explanation, traceability and transfer without the agent’s support.
01 / Understand the system
An answer is not a process.
An agent can organise a sequence of actions, use tools and review the results. Oversight needs to cover that process, not just the quality of its final answer.
- 01
Perceive
Gather the relevant context.
- 02
Plan
Break a goal into steps.
- 03
Act
Use authorised tools.
- 04
Evaluate
Check results and adjust course.
Results return to the context. The cycle continues or stops.
02 / Design before automating
Four questions before choosing a tool.
Start with the learning activity. The autonomy framework proposes setting limits according to the learning goals, the students, the risks and the human review actually available.
- 01
Competence
What must the student not delegate?
- 02
Self-regulation
What support do they need now?
- 03
Risk
What would happen if the agent were wrong?
- 04
Human review
Who can intervene, and how?
Together, these answers define the acceptable scope of autonomy.
03 / The central proposal
Pedagogically Restricted Autonomy.
PRA (APR in the Spanish book) is not a race towards the highest level. It is a set of configurations for matching support to context while protecting the cognitive work through which students learn.
- 00
On-demand assistance
Explicit instructions and direct supervision.
- 01
Support with escalation
Initial support; teacher intervention in assessed work.
- 02
Formative adaptation
Adaptation and feedback with periodic review.
- 03
Bounded pathway
Pathways within defined parameters, alerts and safeguards.
- 04
Bounded autonomy
A narrow domain, traceability and accessible human review.
More autonomy does not mean better learning. Level 4 is exceptional, not a target.
04 / Read the evidence carefully
Not every promise has the same support.
The book distinguishes documented, emerging, plausible and speculative applications. This makes it possible to consider an application without confusing technical feasibility with educational impact.
- 01
Documented
Public evidence establishes the scope of the claim.
- 02
Emerging
An observable implementation; evidence remains limited.
- 03
Plausible
Technically feasible; educational impact still to be tested.
- 04
Speculative
Depends on capabilities not yet reliably available.
Categories must be reviewed as the evidence changes.
A classroom situation · Hypothetical example
A better-written essay does not, on its own, demonstrate better learning.
01
The task
The student must formulate a thesis and defend it with evidence.
02
The permitted support
The agent can point out a gap in the argument and ask for a justification; it does not write the thesis for the student.
03
The check
Afterwards, the student explains their decisions and tackles a related task without the agent’s support.
This activity illustrates a design decision. It is not presented as an experimental case or a guaranteed outcome.
The full map
20 chapters.
A connected argument.
From the concept of an agent to institutional responsibility. Open each part to see its chapters.
IConceptual and ontological foundations
- 01From conversational AI to agentic AI
- 02The ontological problem of the agent: genealogy and contemporary appropriation
- 03Theoretical foundations of agent-mediated learning
IIArchitectures, patterns and coordination
- 04Anatomy and architecture of agentic systems
- 05Design patterns and multi-agent coordination
IIIPedagogical design with agents
- 06Pedagogical roles of educational agents
- 07The Pedagogically Restricted Autonomy framework
- 08Designing pedagogical instructions
- 09Personalisation and adaptation: promises and cautions
IVApplications in higher education
- 10Agents as intelligent tutors
- 11Documented and verified cases
- 12Potential applications across disciplines
- 13Agents for academic research
VReliability, evaluation and governance
- 14Teaching competences for working with agentic AI
- 15Evaluating educational agents: how to know whether they work
- 16Security and systemic reliability
- 17Ethics, bias and educational responsibility
- 18Institutional governance of agentic systems
VIPerspectives and recommendations
- 19Towards a reflective agentic pedagogy
- 20Future scenarios and research directions
VIIAnnexes and supporting material
- AGlossary of concepts and terms
- BDecision matrix for selecting an autonomy level
- CEvaluation template for agentic tools in educational contexts
- DQuestions for institutional reflection before implementing pedagogical agents
- EAgent Learning Canvas: a visual method for designing learning experiences with agents
- FApplication maturity framework: full criteria and classification
- GAnnotated references
Also includes an index of figures. English headings on this website have been normalised for readability. Pagination may change with editorial revisions.
Included with the book
A workbook for moving
from reading to design.
The Agentic AI Teaching Toolkit accompanies the handbook with instruments for preparing, documenting and reviewing an educational activity. It is a workbook, not an AI application.
- 24 pages
- A practical companion, separate from the main book.
- Ten scenarios
- Situations to adapt, not universal recipes.
- Local HTML
- Digital completion in a compatible browser.
- Reading and printing with a fixed layout.
The HTML works offline. Do not enter students’ personal data, and keep copies of your responses; browser storage is not a substitute for a backup.
Before you decide
Understand the scope.
Make your own judgement.
Read the free presentation dossier: cover, chapter-level contents and a summary of the principal frameworks. It does not include the first chapters in full and is not a substitute for the book.
For departments, libraries and teaching units
Need several copies or a shared reading programme?
Ask directly about possible arrangements for a working group, internal development programme or library. Availability and scope are agreed case by case; this page does not promise a fixed institutional package.
Ask the author ↗Before you buy
Do I need to know how to code?
Not to follow the pedagogical argument. The book explains components and architectures; it is not a programming course or a catalogue of prompts to copy.
Which language am I buying?
Each link leads to its own edition: English or Spanish. This website brings both together, but a purchase does not automatically include both languages. Check the contents and final price on Payhip.
Which files are included?
The book and teaching toolkit are offered in PDF and HTML for local reading. Check the Payhip listing for the files included in the edition you intend to buy.
What supports the proposals?
The handbook combines references, case analysis and the author’s own proposals. It distinguishes application maturity and states the limits of its frameworks. A pedagogical proposal is not, on its own, a validated intervention.
Can I read it offline?
The standalone HTML and PDF files are intended for local reading. Opening source links or accessing external services does require an internet connection.
Is the book finished, or will it change?
It is published, with textual revisions planned over the coming weeks. The English edition is still undergoing language review. Content and pagination may change; academic citations should identify the version consulted.
The book + the teaching toolkit
Understand what to delegate.
Know what to preserve.
Choose your reading language. Each edition’s price and included files are shown on Payhip before purchase.
English edition
Language and editorial review in progress · PDF and offline HTML
View English edition ↗
Editorial transparency. Textual corrections and amendments are planned over the coming weeks. In particular, the English edition is still undergoing language and copy-editing review; some corrections may change pagination. Before citing or reviewing, please request the latest revision.