Pedagogy for Networked and Agentic Learning

Connection,Connection, activity, evidence, and revision as design principles for teaching

If knowledge is distributed across people, sources, tools, and systems, tools mediate how a learner thinks, contradictions drive development, and self-models regulate what a learner does next, then teaching cannot simply transmit content or wait for learners to construct it alone. A pedagogy for this situation has a narrower and more demanding job. It should help learners build connections, inspect their own activity, test interpretations against evidence, and keep meaningful control over the models that are built about them, whether those models are their own or institutional. This article sets out what the main pedagogical traditions contribute to that job, what each leaves out, and how they can be combined into a working cycle.

1. Several Traditions, Each With a Limit FoundationalKnowledge that endures for decades — core principles

TheThe traditions below are not mutually exclusive, and most good teaching draws on several at once. Each is given here with what it contributes and where its characteristic limitation lies.

Instructivism contributes clear explanation, explicit modelling of procedures, reduced initial uncertainty, and the reliable communication of foundational knowledge. Where safety depends on getting something right the first time, direct instruction is the appropriate tool. Its limitation is the assumption that knowledge can be transferred intact from the teacher to the learner, so that understanding is what remains after the explanation has been received.

Constructivism contributes the insight that learners interpret new material through what they already believe. Active interpretation, attention to prior knowledge, conceptual change, inquiry, and authentic problems all follow from it. Its limitation is a tendency to place too much weight on individual construction, as though a learner could rebuild a discipline unaided.

Social constructivism builds on Vygotsky's account of development through tools and signs, in which the means of thinking are inherited from a culture and learned through interaction with others[6]. It contributes dialogue, collaboration, scaffolding, participation in disciplinary practice, and shared meaning-making. Its limitation is that agreement within a group does not guarantee accuracy or justice. A community can converge on an error as readily as on a truth.

Constructionism contributes learning through making: public, inspectable artefacts that can be tested, revised, and shared, with reflection occurring through the act of externalising an idea. Its limitation is that making something does not by itself establish that the making was understood, so artefacts need to be read against the reasoning that produced them.

Activity theory treats learning as an activity with a subject, an object, mediating tools, rules, a community, and a division of labour. Yrjö Engeström's extension of Vygotsky's work emphasises that tensions among these elements can push the activity to develop[5]. It explains why adding an AI assistant, a learning log, a rubric, a concept map, or an analytics dashboard changes the whole learning activity rather than one step within it. Its limitation is analytical rather than practical: the framework is demanding to apply, and it describes change more readily than it prescribes a single intervention.

Connectivism contributes the view that knowledge is distributed across networks and that learning involves forming, maintaining, and traversing connections. George Siemens proposed it as a theory suited to the digital age[1], and Stephen Downes developed the relational formulation that knowledge consists of connections formed through experience with a knowing community, with learning as the growth and modification of those connections[2]. It contributes attention to knowing where and whom, to the evaluation and pruning of connections, and to recognising patterns across domains. Its limitation is a tendency to celebrate networks. Popularity is not reliability, and a network can reproduce error, conformity, and inherited exclusion as easily as insight.cf. learning as a network

Critical pedagogy, associated with Paulo Freire, contributes questions about authority, exclusion, and whose knowledge counts[4]. It asks who constructed a network, who is visible within it, who controls ranking, and which connections an institution rewards. It also insists that learners should be able to contest the representations they are given. Its limitation is that critique alone does not tell a learner what to do next, so it needs to be paired with a concrete account of activity.

Metacognition contributes planning, monitoring, calibration, strategic control, reflection, and revision of self-knowledge. John Flavell's original formulation distinguished metacognitive knowledge, metacognitive experiences, tasks, and strategies[3]. Its limitation is that a learner who monitors without being able to change strategy gains little, and a learner who changes strategy without monitoring has no means of judging whether the change helped.

Narrative approaches contribute continuity, the integration of experience, the interpretation of development over time, and the formation of disciplinary and professional identity. Their danger is that retrospective storytelling produces an artificially coherent learner, one whose path looks more direct than it was. A record of what actually happened is a necessary check on that tendency.

2. A Unified Learning Architecture Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheseThese traditions can be combined into a single picture of how learning operates. Sources, people, and tools surround a knowledge network. The network supports learning activity, which produces action, feedback, and evidence. That evidence is recorded in a learning log, which in turn informs the learner's self-model. The self-model shapes future participation, which changes the network, produces new evidence, and revises the model again.the log is not just data it is identity work

                      SOURCES
              research, data, media
                       │
                       ▼
PEOPLE ◄──────── KNOWLEDGE NETWORK ────────► TOOLS
peers, teachers,    concepts and             AI, maps,
communities         relations                repositories
      │                 │                          │
      └─────────────────┼──────────────────────────┘
                        ▼
                 LEARNING ACTIVITY
                        │
             action, feedback, evidence
                        ▼
                   LEARNING LOG
                  recorded history
                        │
                        ▼
                    SELF-MODEL
          capability, strategy, identity
                        │
                        ▼
                FUTURE PARTICIPATION
                  (feeds back into the network,
                   new evidence, and the model)

The diagram is a set of relations rather than a pipeline. Each arrow can run in both directions, and the feedback from future participation is what makes the architecture a learning system rather than a sequence of steps.

3. The Recurring Cycle Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe architecture becomes practical when it is turned into a cycle that a learner or a teacher can run repeatedly. The eight steps below are one such cycle.

1. Orient. Identify the problem, the learner's prior understanding, the points of uncertainty, the relevant people and sources, and the intended outcome.
2. Connect. Find conceptual neighbours, follow references, consult peers or specialists, relate the work to earlier activity, and identify useful tools.
3. Predict. Ask the learner to make part of the current model visible before acting: what do I expect to happen, why, and what evidence would surprise me? This makes later contradiction educationally usable. It is similar in spirit to productive failure, in which learners attempt complex problems before instruction and the failed attempts prepare them to learn from it[7].
4. Act. Produce an artefact, test an idea, run an experiment, explain a concept, attempt a solution, or contribute to a discussion.
5. Record. Capture the intention, the action, the result, the evidence, the difficulty, the decision taken, and the remaining uncertainty.
6. Triangulate. Compare the result with the literature, with feedback, with competing interpretations, and with the wider problem. The triangulation approach used in this site's project guidance is one concrete form of this step, in which a learner repeatedly orients among their own implementation, the existing literature, and the domain problem(see the triangulation page).
7. Reflect. Ask what changed, which strategy helped, which connection misled, what the evidence says about an earlier judgement, whether the present difficulty is a local performance judgement or a global claim about identity, and whether the contradiction belongs to the learner's model or to the activity system around it.
8. Revise. Update the artefact, the concept map, the plan, the source network, the explicit explanation, the self-model, and, where necessary, the curriculum itself.

Two features of the cycle matter. First, it is iterative, so that a failed plan returns the learner to orientation rather than ending the activity. Second, the reflection step is separated from the act step, so that interpretation is informed by the record rather than reconstructed from memory alone. Where a learner is stuck within one mode of work, the unstuck ladder offers a way of moving the activity to another mode while keeping the cycle running.logs beat memory for honest reflection

4. Learning Records as Knowledge Graphs Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA learning record can be represented as a simple graph of typed relations. The representation is useful because it makes the structure of the record inspectable and because it shows how the record connects to later interpretation:

Learner
  ├── attempted ──> Activity
  ├── produced ───> Artefact
  ├── consulted ──> Source
  ├── received ───> Feedback
  ├── revised ────> Interpretation
  └── adopted ────> Strategy

Activity
  ├── occurredAt ──> Time
  ├── concerned ───> Concept
  ├── resultedIn ──> Outcome
  └── evidencedBy ─> Record

It is important to state what this graph is. It is not the learner. It is a selective representation of recorded educational activity, and it inherits every limitation that any selection carries. A self-model may draw on it, following a path from recorded history to a proposed pattern, then to the learner's own inspection, then to an accepted, revised, or rejected interpretation, and finally to a provisional self-model:

Recorded history
      │
      ▼
Proposed pattern
      │
      ▼
Learner inspection
      │
      ▼
Accepted, revised, or rejected interpretation
      │
      ▼
Provisional self-model

The final step is the one most easily skipped. A pattern the system proposes should not become a self-model until the learner has inspected it against their own experience. Without that step, the graph's structure can be mistaken for the learner's identity. The other articles in this series return to this point from the side of learning logs and narrative identity.the map is not the territory

5. Trust in Learning Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TrustTrust operates at several points in a networked learning environment, and each needs to be examined in its own right. A learner has to decide whether to rely on a document, a model, a person, or an institution as a source. They have to decide whether a connection reflects a meaningful relation or only a superficial proximity. They have to ask whether a piece of feedback comes from someone with relevant expertise who understands the learner's objective. They have to check whether their own confidence is calibrated against the evidence. When analytics produce a category or prediction, they need to know which records support it. When an AI assistant offers a reflection, they need to know whether it retrieved evidence, inferred a pattern, or generated a plausible narrative. And they need to know whether a learning log is a contemporaneous record, a later reconstruction, an edited version, or a document written for assessment.

Taken together, these questions show that metacognition is partly a calculus of trust applied reflexively. It involves deciding when to rely on one's memory, one's confidence, one's current strategy, one's interpretation of an event, and one's present model of oneself. The calculus is not a matter of trusting or distrusting oneself in general. It is a matter of matching reliance to the evidence available for a particular claim and a particular consequence. The other articles in this series develop the monitoring and control side of this in detail, and the calculus-of-trust page treats the underlying relational account of trust.the relational account of trust is developed there

6. What the Cycle Depends On Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe cycle is not self-running, and four conditions in particular determine whether it produces learning rather than activity. The first is that the connections a learner builds are chosen with some awareness of their quality, so that a large network does not substitute for a reliable one. The second is that the learner is able to inspect their own activity, which requires a record that can be compared against later interpretations. The third is that feedback is informative, meaning it is specific enough to change the next action. The fourth is that the institution remains accountable for the environment in which the cycle runs. An individual learner cannot repair a curriculum that makes the predict and reflect steps impossible, for example by allowing no time for false starts. The traditions above each draw attention to one of these conditions, and a pedagogy that attends to only one of them will be incomplete.

Related Topics

• Pedagogy, Andragogy, Heutagogy, Rhizome: Who Decides What You Learn — the sequence of answers to who decides what is learned, which sections 1 and 2 extend.
• How Much Structure to Impose: A Design Decision Guide for Educators — how to set the balance between structure and learner direction in a design.
• Personal Learning Maturity — the levels of self-directed learning that the cycle in section 3 assumes a learner is progressing through.
• Learning by Teaching, Learning by Building — two ways of exposing what a learner does not yet understand, which the predict and act steps make use of.
• Project Navigation and the Art of Triangulation — the triangulation step in section 3 in its project-guidance form.
• The Unstuck Ladder — switching modes of work when one stops progressing.
• Words as Tools: Meaning, Mediation, and Activity — the account of mediation and activity that sections 1 and 2 draw on.

References

1. G. Siemens, "Connectivism: A Learning Theory for the Digital Age," International Journal of Instructional Technology and Distance Learning 2(1), 2005.
2. S. Downes, "Learning Networks and Connective Knowledge," IT Forum, 2006. https://philarchive.org/archive/DOWLNAv1
3. J. H. Flavell, "Metacognition and Cognitive Monitoring: A New Area of Cognitive-Developmental Inquiry," American Psychologist 34(10), 1979, pp. 906–911.
4. P. Freire, Pedagogy of the Oppressed, trans. M. B. Ramos, Herder and Herder, 1970.
5. Y. Engeström, "Expansive Learning at Work: Toward an Activity Theoretical Reconceptualization," Journal of Education and Work 14(1), 2001, pp. 133–156. https://doi.org/10.1080/13639080020028747
6. L. S. Vygotsky, Mind in Society: The Development of Higher Psychological Processes, eds. M. Cole, V. John-Steiner, S. Scribner, and E. Souberman, Harvard University Press, 1978.
7. M. Kapur, "Productive Failure," Cognition and Instruction 26(3), 2008, pp. 379–424.