Share
Embed
Transcript
46 slides
-
The arc of this talk 01 Debt is the price
of learning 02 Be in control, and in learning 03 The ways we know -
1. Debt is the price of learning What changed when
AI accelerated execution? GOTO Copenhagen -
Debt as a business strategy “Shipping first time code is
like going into debt. A little debt speeds development so long as it is paid back promptly with a rewrite. Ward Cunningham OOPSLA 1992 The WyCash Portfolio Management System Every minute spent on not-quite-right code counts as interest on that debt.” -
Debt as a learning strategy “I am never in favor
of writing code poorly, but I am in favor of writing code to reflect your current understanding of a problem even if that understanding is partial. Ward Cunningham 2011 Debt Metaphor I thought that rushing software out the door to get some experience with it was a good idea, but you would eventually go back and … put that learning back into the program.” -
Debt can be rational when acting now creates the learning
needed to improve later. -
Debt metaphor as a communication strategy Design flaw is like
reckless debt Business strategy Martin Fowler 2014 Technical Debt Quadrant Learning strategy -
Every feature eliminates some future optionality It costs us to
maintain or delete the feature. Kent Beck 2026 Earn *AND* Learn -
Technical debt is learning made visible in software UNDERSTANDING₁ →
SOFTWARE₁ → EXPERIENCE → UNDERSTANDING₂ Software₁ may no longer fit what we now understand. Yesterday’s good understanding can become today’s debt. Code as fossil records -
Debt is the price of Learning. What changed when AI
accelerated execution? -
Earlier, code could only move as fast as developer understanding
EXECUTION · PRODUCTION Spec → Code → Features ↑↑↑ UNDERSTANDING · ACCOUNTABILITY ↑↑↑ -
AI changes the economics Delivery speed is no longer a
good proxy for measuring learning debt. EXECUTION · PRODUCTION Spec → Code → Features ↑↑↑ UNDERSTANDING · ACCOUNTABILITY ? -
AI exposes the real bottleneck by removing a constraint BEFORE
MAKE UNDERSTAND MAKE Historically, lack of understanding often slowed production because humans were doing the production. WITH AI MAKE MAKE We can keep producing despite weak shared understanding. Code becomes fossil records of the agents’ understanding. MAKE UNDERSTAND? -
Externality: Local learning debt becomes social debt A tragedy of
the commons - accelerated debt creation and interest LOCAL BENEFIT COLLECTIVE COST Generate Ship Move on Review · Comprehend Integrate · Maintain Trust Making outruns making sense together -
What happens when ‘our’ capacity to produce exceeds our capacity
to understand what’s produced? How do we get back in control? -
2. Be in control, and in learning Which loop are
we in? GOTO Copenhagen -
We must pace the frontier Pay the learning debt -
“Understand and control these systems” (Essay, Sept 12, 2026) “AI has been advancing drastically faster, driven primarily by AI’s growing ability to build the next generation of AI. This dynamic is called recursive self-improvement… Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all. “ -
A timely response: improve the execution loop Harness engineering ensures
fidelity between intent and execution Credit: Birgitta Böckeler -
Spec-Driven Development as part of the harness Specifications as the
primary source of truth INTENT SPEC CONTEXT AGENT TESTS Functional & Non-functional SDD makes intent explicit before asking the agent to execute VERIFY -
DDD-informed SDD Many similar experimentations in our community “ESML (Event
Storming Markup Language) is a faithful, high-fidelity transcription of what's on the board — stickies, colors, positions, groupings — without interpretation. DMML (Domain Model Markup Language) interprets the ESML into a strategic and tactical DDD domain model — subdomains, bounded contexts, aggregates, entities, value objects, events, commands, policies, and more”. Credit: Joseph Yoder & Marden Neubert -
A harness is a frame INTENT SPEC CONTEXT AGENT TESTS
What if the frame needs to change? VERIFY -
A harness can only go so far Human experience and
intuition remain part of the “implicit harness” SITUATED JUDGMENT What “good” looks like in this context Birgitta Böckeler 2026 Harness Engineering for coding agent users “Harnesses are an attempt to externalise and make explicit what human developer experience brings to the table, but it can only go so far.” ORGANISATIONAL ALIGNMENT What debt is tolerated, and why ORGANISATIONAL MEMORY We don’t do it that way here SOCIAL ACCOUNTABILITY What this means for people maintaining the code -
Double-loop learning LEARNING LOOP Assumptions Spec Execution Why we do
what we do What we want to do How do we do it (“Theory Building”) EXECUTION LOOP Credit: Chris Argyris & Donald Schön Organizational Learning II: Theory, Method, and Practice. (1996) -
Better specs improve execution with a frame. But sometimes we
need a new frame. LEARNING LOOP Assumptions Spec Execution Why we do what we do What we want to do How do we do it (“Theory Building”) EXECUTION LOOP Credit: Chris Argyris & Donald Schön Organizational Learning II: Theory, Method, and Practice. (1996) -
What if surprise is a feature? “The XP goal is
not to have results match expectations. Kent Beck 2026 The Third Loop Podcast I prefer to have results not match expectations in positive ways.” -
Keep execution and learning in sync Execute on current understanding.
Refactor when understanding changes. EXECUTION LOOP LEARNING LOOP DDD / XP / COMO SDD / HARNESS Externalize intent Govern execution “How do we align output with specs?” ⇄ Challenge assumptions Reframe intent “Why are we modeling it this way?” -
Execution Loop. Learning Loop. Different mindset. CONTROL MINDSET TECHNICAL EXECUTION
Intent → Spec → Harness → Software → Outcome → Learning deviation = problem Did the outcome match our expectations? SOCIOTECHNICAL LEARNING Intent ⇄ People ⇄ Conversations ⇄ Models ⇄ Software ⇄ Learning Which assumptions produced the current intent? Do we still want the same thing? LEARNING MINDSET deviation = information -
Learning is the point Are the specs optimized for human
learning? Code is a side effect “Software development is a learning process. Working code is a side effect.” Alberto Brandolini Credit: Alberto Brandolini “Code is still a side effect” (2026) -
How do we think across different loops? How do we
know when our ways of knowing needs to change? -
3. The ways we know What actually happens when we
think and learn together? GOTO Copenhagen -
The 4 ways of knowing A language for knowing beyond
propositions PROPOSITIONAL logic · semantic · mental model · beliefs ↑ PROCEDURAL know-how · skills · power ↑ PERSPECTIVAL embodied · presence · situated · episodic ↑ PARTICIPATORY agent–arena coupling · meaning · identity Credit: John Vervaeke -
Relevance realization How do we know what to pay attention
to? SALIENT What matters here and now, in this state of mind ? Intelligence includes discovering what is worth noticing. Credit: John Vervaeke -
Relevance realization How do we know what calls for action?
POSSIBLE → RELEVANT → SALIENT → ACTIONABLE Intelligence includes discovering what deserves attention, and action. Credit: John Vervaeke -
Have you noticed … When WE use mental shortcuts, we
call them heuristics. When OTHERS use mental shortcuts, we call them biases. What feels relevant depends on where we stand. Relevance realization contains the risk of “self-deception”. Credit: John Vervaeke -
Dialogic relevance realization We need each other to escape self-deception
and update mental models. IN A DIALOGUE MY PERSPECTIVE DIALOGIC SPACE YOUR PERSPECTIVE Your reaction can make my assumption visible. My question can change what becomes relevant to you. Together, we can discover what neither of us could see alone. A dialogue affords reciprocal opening, not just information exchange. Credit: David Bohm, John Vervaeke, Rupert Wegerif -
A dialogic practice for reciprocal opening Notice → Welcome →
Attend → Share → Listen → Reflect → (Possible) Insight Perspective Taking Opponent Processing Reciprocal opening From reaction to curiosity about my own assumptions. Different perspectives make our assumptions visible. A new frame can emerge that none of us saw at the start. Dialogue does not manufacture insight. It creates conditions in which reframing can emerge. -
The frame itself becomes available for inquiry DDD PERSON’S MENTAL
MODEL SDD PERSON’S MENTAL MODEL Detailed specs constrain learning Detailed specs align execution with intent explicitness-through-action explicitness-before-action I encounter difference → I experience discomfort → I welcome my reaction → my tacit assumption becomes visible → I share my perspective → I look from the other’s lens → our perspectives constrain and afford each other → my relevance landscape reorganizes → We both see differently Credit: David Bohm, John Vervaeke -
Dialogue as a sociotechnical design tool Design goal: Perspectival flow
and reciprocal opening PERSPECTIVAL embodied · presence · situated · episodic PARTICIPATORY agent–arena coupling · meaning · identity Dialogue: update the perspectival through the participatory Credit: John Vervaeke -
Modeling together can change the modelers DDD PERSON’S MENTAL MODEL
SDD PERSON’S MENTAL MODEL Detailed specs constrain learning Detailed specs align execution with intent explicitness-through-action explicitness-before-action What if we use SDD to externalize the current state of understanding, and use DDD to continuously challenge and evolve that understanding? explicitness-before-action AND explicitness-through-action Credit: David Bohm, John Vervaeke -
Agile was almost “Conversational” At Snowbird, Kent Beck proposed “Conversational
Software Development” rather than agile “Sitting together and working on something is such a powerful source of energy, feedback and value, that I want to find ways to bring people together, not ways to Kent Beck 2026 The Third Loop Podcast successfully keep people apart.” -
Bohmian Dialogue “A dialogue must have no other goal but
to be able to talk together. A group starts as a microcosm of the whole culture. We want to be part of the group but we are also afraid of losing our individuality. David Bohm (1917-1992) American Physicist A dialogue loosens the collective sociocultural rigidity that holds all of us in its grip. In a dialogue we suspend assumptions and become witness to the limits and boundaries of our own thoughts. ” -
What if coupling is the point? Coupling is not a
liability among humans. Coupling is a capacity for the unknown. SOFTWARE OPTIONALITY SOCIAL OPTIONALITY Can the software respond to change? Can we respond together to change? The Future is not something we specify. It is something we remain capable of responding to. Credit: John Vervaeke, Kent Beck