Fragments: February 13

Martin Fowler · article · 2026-02-13 source ↗ Reader 🖥️

Agent-Native Design AI Labor Disruption Burnout & Recovery Cognitive Debt Developer Experience & Flow The Genie (Beck)Human as Orchestrator of Agents Orchestration Tax / Human Review Bottleneck Pair & Mob Programming Supervisory Programming Technical Debt

As LLMs become established coding collaborators, software development is reorganizing around 'supervisory programming' in which developers herd fleets of agents — a shift that recasts the profession's central challenges as preserving a team's shared understanding of the system (cognitive debt), sustaining a developer experience that is effectively identical to agent experience, and containing the work-intensification and context-switching that driving many agents creates.

mixed · mixed — A set of loosely connected fragments — field notes from the Thoughtworks Future of Software Development Retreat plus reactions to others' posts. Mostly expository (reporting attendees' observations) with argumentative interludes (Fowler's cognitive-debt reframing, his inclination on team size) and forward-looking conjecture.

Structure

Part	Role
Roles of senior developers	Argues seniors retain a bright future by shifting toward architecture, and that managing LLM agents resembles managing junior developers; adds that hands-on experience is what converts skeptics
Junior vs mid-level vs senior developers	Reassigns the anxiety — juniors will still be needed and can use LLMs as an always-available mentor; mid-level developers face the greatest challenge
Cognitive debt (Storey) and the debt metaphor	Introduces cognitive debt via Storey's post and reframes it through Fowler's technical-debt lens: cruft (= ignorance) vs. the debt metaphor for reasoning about its costs
DevEx = Agent Experience (Tacho)	Argues that the smooth tooling, clear environment, modularity, and naming that serve developers serve LLMs equally — the two Venn circles coincide
IDEs incorporating LLMs	Argues IDEs must blend deterministic features with LLMs — using LLMs to orchestrate the IDE's refactoring rather than editing the whole codebase, and to help users choose the right tool
Team size and pair programming	Conjectures that teams stay the same size and do more (two-pizza balance) rather than shrinking, and that two humans driving agents may beat one
Work intensification (HBR) and the manager's fatigue (Fournier)	Surfaces the cost side: AI adoption intensifies work toward burnout, and 'everyone becomes a manager' brings context-switching fatigue; closes on supervisory programming and accountability

Key terms

cognitive debt Cognitive Debt
The accumulating cost of a team losing its shared 'theory of the system' — the understanding of why design decisions were made and how parts fit together; Fowler maps it onto his technical-debt metaphor, where the cognitive equivalent of cruft is ignorance of the code and its domain
cruft
The bad stuff itself that accumulates in a codebase (poor module boundaries, bad naming); Fowler distinguishes it from the debt metaphor, which is his tool for reasoning about the costs cruft imposes — cognitively, the equivalent of cruft is ignorance
The Genie The Genie (Beck)
Fowler's (Beck-derived) name for the LLM as a code-generating collaborator; invoked as a party to whom the same costs and legibility benefits apply as to humans ('The Genie's Galaxy Brain', 'The Genies')
Developer Experience = Agent Experience
The thesis, credited to Laura Tacho, that the practices that improve developer experience also improve LLM effectiveness — 'The Venn Diagram of Developer Experience and Agent Experience is a circle'
supervisory programming / supervisory programmer
The anticipated shift from programmers engaged directly with the code to programmers herding a bunch of agents while remaining accountable for the code generated under their watch
two-pizza team
The team-size heuristic Fowler invokes to argue teams will not shrink to 'one-pizza' size just because 'LLMs don't eat pizza'; the size balances collaboration benefits against coordination costs
so January
A quip that outdated dismissals of LLM capability are already stale — 'some negative opinions of LLM capabilities are so January' — given how fast models have improved

Propositions

explicitSenior developers retain a valuable future by shifting toward architectural issues rather than the messy details of syntax and coding as LLMs absorb lower-level work.conf: high
explicitManaging LLM agents has strong similarities to managing junior developers.conf: medium
explicitcorroboratedHands-on practical experience is what converts LLM-skeptical senior developers, because it gives them credible information to judge the value.conf: medium
supports ←AI Enthusiasts Are in a Race Against Time, AI Skeptics Are in a Race Against Entropy · p12 — Her prescription that skeptics build real AI expertise to earn standing rests on the empirical mechanism that hands-on practical experience is what converts skeptics by giving them credible grounds to judge value.
explicitcorroboratedMid-level developers, more than juniors or seniors, face the greatest challenge in the LLM era, because they formed their careers without LLMs but lack the experience to drive them as effectively as seniors.conf: medium
supports ←The case for making your own apps · p6 — Mann's independent claim that AI benefits the very junior and very senior most, leaving the middle in a 'weird spot,' converges with Fowler's mid-level-hardest-hit observation, adding Mann's specific mechanism (juniors: untapped-intelligence uplift; seniors: an agent 'farm' to steer).
explicitLLMs can serve junior developers as an always-available mentor, though juniors should keep a certain skepticism of AI mentors — as they should of human ones.conf: medium
explicitCognitive debt — the fragmentation or loss of a team's shared theory of the system — can paralyze a team faster than technical debt.conf: high
explicitcorroboratedThe debt metaphor extends to the cognitive realm: the cognitive equivalent of cruft is ignorance of the code and its domain, and a team either pays interest (each change harder) or pays down principal by explicit investment in knowledge — a choice that depends on relative costs, and that now applies to both humans and the Genie.conf: high
supports ←Technical Debt Quadrant · p7 — This is Fowler, sixteen years later, doing exactly what this proposition worries about being hard to do: extending the interest/pay-down-principal logic from technical debt to cognitive debt ('the debt metaphor extends to the cognitive realm... a team either pays interest... or pays down principal'). The original quadrant piece supplies the financial mechanics that the later fragments post explicitly reuses for a domain the 2009 piece itself flagged as awkward to extend the metaphor into.
explicitDeveloper Experience and Agent Experience are effectively the same thing — the smooth tooling, clear environment information, good modularity, and descriptive naming that help developers also help LLMs generate code quickly and correctly.conf: high
explicitIt is notable and regrettable that executives will invest in smoothing the development path for LLMs while declining to make the equivalent effort for human developers.conf: medium
explicitIDEs should use LLMs only for tasks that resist deterministic methods and otherwise have the LLM orchestrate the IDE's deterministic refactoring capabilities, rather than operating an LLM over the entire codebase for mechanical changes like a clustered rename.conf: medium
explicitcontestedDevelopment teams will likely stay roughly the same size and accomplish much more, rather than shrinking, because the two-pizza size effectively balances the benefits of human collaboration against the costs of coordination.conf: medium
contradicts ←Gene Kim on Vibe Coding - Part II | Episode 411 | Troubleshooting Agile · p1 — Kim argues vibe coding grows team size/headcount via latent demand; Fowler's fragments argues two-pizza teams stay roughly the same size and simply accomplish more — both reject shrinkage, but for different reasons (growth vs. stasis).
explicitTwo humans driving a bunch of agents may outperform a single programmer driving agents, combining the benefits of pairing with the greater code-generative ability of the LLMs.conf: medium
explicitEnthusiastic AI adoption tends to intensify work — faster pace, broader scope of tasks, and longer hours, often unbidden — which can become unsustainable and lead to cognitive fatigue, burnout, and lower-quality output.conf: high
explicitdevelopedAs programmers become supervisors of many agents they remain accountable for the code generated under their watch, and the hardest emerging cost is context-switching — so effective supervisory-programming practice must harvest the parallelism of agents while minimizing that context-switching.conf: high
extends ←Software Teaming in the Age of AI with James Herr and Woody Zuill · p5 — Fowler's claim names context-switching as the hardest emerging cost of supervising many agents and calls for practice that 'harvests parallelism while minimizing context-switching'; 'plate spinning' is a concrete team-level technique aimed at exactly that, though this episode doesn't verify it actually reduces (rather than redistributes) the switching cost — see gap 3.

Arguments

Loss of shared system understanding should be managed deliberately as a form of debt — cognitive debt — analogous to but distinct from technical debt. strength: medium

Explicit premises:

Storey's student teams could no longer make simple changes by weeks 7-8, and the real problem was not messy code but that no one could explain why design decisions had been made or how parts fit together — their shared theory of the system had fragmented.
In Fowler's technical-debt framing, 'cruft' is the bad stuff itself, while the debt metaphor is a way to reason about the costs cruft imposes: pay interest or pay down principal.
The cognitive equivalent of cruft is ignorance of the code and the domain it supports.

Implicit premises:

A team's ability to change software depends on shared understanding, not only on code quality.
The interest-versus-principal cost logic that works for technical debt transfers cleanly to knowledge.

Inferences:

Therefore one either accepts higher ongoing costs (harder future changes) or makes an explicit investment to gain knowledge, choosing by relative cost — and these costs now fall on both humans and the Genie.

Investment in developer experience is simultaneously investment in agent effectiveness, so the two are worth funding together. strength: medium

Explicit premises:

Tacho's claim that the Venn diagram of Developer Experience and Agent Experience is a circle.
Smooth tooling and clear information about the development environment help LLMs figure out how to create code quickly and correctly.
There is growing evidence that good modularity and descriptive naming are as good for the transformer as for human ('squishy') neural networks.

Implicit premises:

LLMs and humans face structurally similar comprehension bottlenecks when navigating a codebase.
Even a high-capability model (the Genie's 'Galaxy Brain') still benefits from a legible codebase rather than merely tolerating a confusing one.

Inferences:

Therefore the same practices serve both audiences and DevEx work is dual-purpose — though it is telling that execs will make this effort for the robots but not for the humans.

LLMs will not shrink development teams; teams of roughly the same size will do much more. strength: low

Explicit premises:

The two-pizza team size effectively balances the benefits of human collaboration against the costs of coordination.
LLMs 'don't eat pizza' — agents add capacity without adding to the human coordination unit.

Implicit premises:

The binding constraint on human team size is coordination cost, not the volume of code that can be produced.
Adding agent capacity raises throughput without changing the human coordination calculus.

Inferences:

Therefore the equilibrium human team size is stable and the added agent capacity increases output rather than reducing headcount.

Narratives

In an entrepreneurship course, student teams shipped features fast but hit a wall by weeks 7-8, unable to make even simple changes without breaking things; they blamed technical debt, but the real problem was that no one could explain why decisions had been made or how the parts fit — their shared theory of the system had fragmented. They had accumulated cognitive debt faster than technical debt, and it paralyzed them.
→ Concrete illustration that grounds the cognitive-debt concept and motivates Fowler's debt-metaphor reframing
Senior developers who were very resistant to using LLMs were put through an exercise that forced hands-on work with them; a third were instantly converted to being very pro-LLM. As was quipped, some negative opinions of LLM capabilities 'are so January'.
→ Evidence that experience, not argument, shifts skeptics — and that model capabilities move fast enough to date opinions within weeks
A domain rename of 'person' to 'contact' spans function names, field names, documentation, and test cases, so a simple search-replace is not enough; rather than run an LLM over the whole codebase, the LLM chooses to orchestrate the IDE's refactoring capabilities on each site it finds — and an attendee noted that renames analyzed in an IDE occur in clusters like this.
→ Concretizes the thesis that LLMs should choreograph deterministic IDE features rather than replace them
An eight-month study of a ~200-employee U.S. technology company found that generative AI led employees to work at a faster pace, take on a broader scope of tasks, and extend work into more hours of the day, often unbidden; once the novelty faded, workload creep drove cognitive fatigue, burnout, and lower-quality work.
→ Evidence for the work-intensification cost of AI adoption and a counterweight to the productivity-surge optimism

Questions addressed

✅ What role do senior developers play as LLMs become established?
They retain a bright future by focusing more on architectural issues and by managing LLM agents much as they would manage junior developers.
✅ Will junior developers be replaced by untiring agents?
No — juniors will still be needed (partly for their open-minded LLM fluency) and can use LLMs as a mentor; it is mid-level developers who face the greatest challenge.
✅ How should we think about the loss of a team's shared understanding of its system?
As cognitive debt, reasoned about with the technical-debt metaphor: cruft is ignorance of code and domain, and one pays interest or pays down principal by investing in knowledge.
✅ Should organizations still invest in Developer Experience given LLMs?
Yes — because Developer Experience and Agent Experience coincide, the same investments benefit both humans and LLMs.
⭕ What will programming work look like as agents proliferate?
A shift toward supervisory programming — herding many agents while remaining accountable for their code — with context-switching as the key emerging problem and effective workflows still to be discovered.

Gaps

incompleteThe fragments assert that mid-level developers face the greatest challenge but offer no guidance on what mid-level developers should actually do to adapt or reskill.
Read next: Seek writing on reskilling paths and career development for developers in the AI era, and on how mid-career practitioners build LLM-driving experience.
closed byThe third golden age of software engineering – thanks to AI, with Grady Booch — Booch offers concrete guidance for displaced/mid-level developers: move up a level of abstraction toward systems-level thinking, and study specific foundational texts (Simon & Newell, Minsky's Society of Mind) rather than leaving 'what to do' unspecified.
incompleteCognitive debt is named and mapped onto the debt metaphor, but there is no method for detecting, measuring, or paying it down concretely — and the technical-debt practices (refactoring, restructuring) do not obviously transfer to restoring a shared theory of the system.
Read next: Read Storey's full post, plus work on knowledge bases, ADRs, and documentation practices aimed at preserving shared understanding across humans and agents.
closed byTechnology Radar, Volume 34 — The Radar's 'codebase cognitive debt' blip supplies the missing 'method for detecting, measuring, or paying it down' that the gap asked for — naming feedback sensors for coding agents, tracking team cognitive load, and architectural fitness functions as concrete countermeasures — though still short of a worked, measurable procedure.
uninformedThe claim that good modularity and descriptive naming help the transformer 'as much as' human developers rests on unspecified 'growing evidence' that is neither cited nor quantified, leaving the strength of the DevEx=AgentEx equivalence unverified.
Read next: Look for empirical studies measuring how code structure, modularity, and naming affect LLM code-generation accuracy versus human comprehension.
illogicalThe piece holds that teams will stay the same size and 'do much more' while, a few fragments later, citing an HBR study showing that doing much more (faster pace, broader scope, longer hours) is precisely what drives cognitive fatigue and burnout — a tension it never reconciles, and for which it proposes no countermeasures.
Read next: Read on sustainable pace, workload governance, and DevEx/wellbeing metrics under AI, and on how 'same-size team, much more output' is kept humane.

Fragments: February 13

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents