Fragments: July 6

Martin Fowler · article · 2026-07-06 source ↗ Reader 🖥️

Agent Harness Agentic Engineering AI as Amplifier AI Dependence Risk AI Unit Economics Changeability & Optionality Code Quality The Genie (Beck)Mechanical Sympathy (for LLMs)Software Design Fundamentals Token Efficiency Usage Limits & Rationing

Agentic software development has crossed from aspiration into established practice — the debate is no longer whether it changes software engineering but how — and the central live question is whether classic design and architecture fundamentals still matter, with Fowler's emerging answer being that they do, both because agents use the same constructs humans do and because good design hedges against escalating token costs and the risk of losing affordable access to AI.

mixed · mixed — A curated 'Fragments' post: reportage from the second Future of Software Development Retreat plus commentary on several linked pieces. Mixes narrative reportage, expository summary of others' writing, and light argument; explicitly 'not an event that forms conclusions.'

Structure

Part	Role
Retreat framing and the speculation-to-practice shift (Edwards-Alexander, Herlein)	Reports the mood change between the February/Utah retreat and the European one: from hesitancy to confidence that 'the evidence is in', from aspiration to shipping production
Micro-observations: 'harness engineering' as a new term; token-cost anxiety	Two field notes showing how fast the discipline moves and how the concern has flipped from incentivizing any AI use to worrying about token cost
Does architecture and design still matter? (two hypotheses + session tidbits)	Frames the core dispute — the model's 'Galaxy Brain' makes design irrelevant vs. developer experience equals agent experience — and lists practical observations from a session on it
Verraes: design as a hedge against AI-dependence risk	Adds a further reason to value good design: uncertainty about whether AI tools will remain available or affordable
Charity Majors: the ethics of working with AI	Endorses the position that there is no ethical gain from renouncing AI; engagement beats purity
Gergely Orosz: fraudulent DMCA takedown	Recounts a bogus copyright claim erasing an article from Google search and the hoped-for Streisand effect
404media: the token 'tokenpocalypse'	Summarizes leaked evidence that enterprise token costs are exploding and firms are throttling AI use
Declaration of Independence / reading history (Devereaux)	Reflective coda on reading a short but explosive foundational text in its historical context

Key terms

The Genie The Genie (Beck)
Fowler and Kent Beck's metaphor for the LLM — a powerful, capricious agent you converse with to get work done
harness engineering Agent Harness
The (newly named) discipline of building the scaffolding around models; notable here for how fast it arose — 'wasn't even a term in Utah'
mechanical sympathy Mechanical Sympathy (for LLMs)
A cultivated, working understanding of how current LLMs actually operate — used instead of speculating about future capabilities — so you can use them well
Galaxy Brain
The hypothesis that the model is so capable it can absorb any amount of spaghetti, so we no longer need to care about architecture or design
the Venn Diagram of Developer Experience and Agent Experience is a circle
Laura Tacho's claim that what helps humans understand a codebase (modularity, naming) helps agents equally, because both use the same constructs
ease of change
The reduction of internal design quality to how cheaply the code can be changed; proposed here as measurable via token cost
tokenpocalypse
The emerging crisis of runaway enterprise token costs that is driving firms to throttle or ration AI use

Propositions

explicitThe debate over whether agentic development changes software engineering is over; practitioners have moved from arguing 'whether' to arguing 'how'.conf: high
explicitAgentic development shifted from aspirational speculation at the February/Utah retreat to shipping production work by the European retreat.conf: high
explicitcorroboratedThe field moves so fast that 'harness engineering' became a widely-discussed term within months, though it was not even a term at the earlier Utah retreat.conf: high
supports ←Fragments: July 21 · p1 — Restates the same 'harness engineering as emerging discipline' observation, now formalized as a named headline finding in the full Thoughtworks retreat report — though this post never elaborates the term itself.
explicitdevelopedPractitioners' concern has flipped from incentivizing any LLM use to worrying about the cost of tokens.conf: high
extends ←Fragments: July 21 · p10 — Connects board-level cost-cutting enthusiasm for AI to the token-cost worry Fowler flagged two weeks earlier in the same fragments series; he explicitly hopes rising token-cost awareness will temper board eagerness.
explicitWhether architecture and design still matter in the agentic age is contested between two hypotheses: that a sufficiently capable model makes design irrelevant, versus that developer experience and agent experience are the same thing so good design helps agents as much as humans.conf: high
explicitdevelopedInternal design quality reduces to ease of change, and whether that lesson carries over to agents is an open empirical question.conf: medium
extends ←Does code quality still matter? · p7 — Fowler notes that 'design quality reduces to ease of change' and flags whether that lesson carries over to agents as an open empirical question; Seemann's LLaMe thought experiment works through exactly this transfer, arguing ease-of-change-for-LLMs could survive even where human-readability-based quality does not.
extends ←The Economic Benefit of Refactoring · p2 — july-6#p6 flags 'whether the ease-of-change lesson carries over to agents' as an open empirical question; this article directly measures ease-of-change as agent token cost, before and after refactoring the identical codebase, and reports a specific number.
explicitcorroboratedToken cost can serve as a proxy metric for architectural quality: if the same change requires fewer tokens, the architecture is better.conf: medium
supports ←Does code quality still matter? · p8 — Fowler proposes token cost as a proxy for architectural quality; Seemann independently arrives at the same proxy ('more resources (tokens?) for LLMs to understand') when speculating about quality in a hypothetical LLM-native language.
refines ←Downstream #1: Where Do Our Tokens Go? · p3 — Fowler proposes raw token count as an architecture-quality proxy; Bai et al.'s finding (cited here) that identical model+task runs vary up to 30x in cost due to harness and randomness shows the proxy is invalid unless harness and stochastic variance are held constant first — this closes that document's own flagged gap (g1).
extends ←The Economic Benefit of Refactoring · p1 — july-6#p7 asserts token cost as an unvalidated proxy for architectural quality ('if the same change requires fewer tokens, the architecture is better'); this article is the first concrete, named quantitative demonstration of exactly that claim — 83% fewer input tokens after refactoring the same code.
explicitGood architecture reveals its value only over time and cannot be reliably measured in the short term.conf: medium
explicitThe agentic age is the first time in history that the computer itself 'cares about' code quality.conf: medium
explicitdevelopedDevelopers should build mechanical sympathy for how current LLMs actually work rather than speculate about what future models may do.conf: high
extends ←Who is LLM? · p6 — This piece's practical lesson (don't be fooled by lucky 'jackpot' results; real skill takes deliberate learning) is generalized in Fowler's own later Fragments piece (published the same month) into the explicit discipline of building 'mechanical sympathy' for how current LLMs actually work rather than speculating.
explicitcorroboratedLLMs imitate and amplify the existing code they see, so pre-existing design problems get magnified rather than corrected — even with a good harness.conf: high
supports ←Software Design in the Agentic Age: Placing Your Bets · p1 — Same claim (LLMs imitate and amplify existing code, magnifying rather than correcting design problems) independently reported from a different write-up of the same retreat.
explicitcorroboratedGood, human-legible design is a hedge against the risk of losing affordable or permitted access to AI tools (governments blocking models, popular opposition, regulation, rising costs).conf: medium
supports ←Stranded in the Slow Zone · p1 — This is the concrete real-world case of exactly the 'governments blocking models' scenario Fowler names abstractly as a reason good, human-legible design functions as a hedge.
explicitThere is no ethical gain from renouncing AI and condemning its users; engaged, responsible use does more for those harmed than abstention for the sake of purity.conf: high
implicitThe classic design fundamentals that historically served human maintainers are being re-justified on new grounds — token economy and agent comprehension — rather than replaced.conf: medium

Arguments

Classic software design and architecture fundamentals remain important in the agentic age. strength: medium

Explicit premises:

Agents use the same constructs humans do to understand a codebase, so modularity and naming help them as much as us (Tacho, Tornhill).
A better-architected change costs fewer tokens, making token cost a signal of design quality.
Good design is a hedge against the risk of losing affordable or permitted access to AI (Verraes).

Implicit premises:

The historical link between internal quality and ease-of-change persists for agent-driven change.
Token cost is a valid proxy for architectural quality despite confounds.

Inferences:

Therefore design investment is justified even when a highly capable model could brute-force poor structure.

Developers should cultivate mechanical sympathy for how LLMs work rather than speculate about their future capabilities. strength: medium

Explicit premises:

The field moves so fast that yesterday's non-terms (harness engineering) are today's central topics.
Speculating about what LLMs 'may do' in the future is unreliable.

Implicit premises:

Understanding current mechanics yields more actionable guidance than betting on the trajectory.
Better usage of today's models compounds value now.

Inferences:

Therefore effort is better spent understanding how models work than predicting where they will go.

There is no ethical case for renouncing AI use. strength: medium

Explicit premises:

AI has real harms in both training (stolen data) and inference (slop, lack of accountability, skill atrophy).
Abstention and purity provide little practical help to those being harmed.
Showing up and building with the technology does more good than removing oneself from the system.

Implicit premises:

Individual abstention does not materially reduce the harms.
The technology is too powerful and useful to be displaced by refusal.

Inferences:

Therefore ethical energy is better channeled into engaged, responsible use than into purity or condemnation.

Narratives

The retreat trajectory: at Deer Valley there was 'hesitancy and a belief that there was something here even if we weren't yet sure what it was'; at Engelberg (Europe) there was confidence that 'the value is here... the evidence is in.'
→ Dramatizes the shift from speculation to established practice — 'not a conference for true believers'
A concrete agentic workflow: take a story from the backlog, talk it over with an agent, capture the agreement as an ADR for a persistent spec record, generate a task list, and have the agent complete it.
→ Illustrates an emerging repeatable practice pattern for agent-driven development
The 'tokenpocalypse': 404media's leaked Slack chats, dashboards and emails from firms including Citi and Amazon show one company's token bill rising from $5M (Aug 2025) to $15M (May 2026), on track for over $120M/year; Accenture found the biggest driver was staff 'chewing tokens' on things like turning PDFs into slides, not agentic coding; one mitigation makes Claude/Codex 'talk like cavemen'.
→ Concrete evidence that token cost has become a first-order operational constraint driving throttling and rationing
Gergely Orosz's Pollen-collapse article was erased from Google search by a bogus DMCA claim filed by 'Ellie Piee' (listed as living on the uninhabited Bouvet Island), falsely alleging it copied an unrelated New York Post piece 'Band Leader Hits Winning Chord'; Fowler hopes the Streisand effect bites those responsible.
→ Standalone cautionary tale on the fragility of online information and weaponized copyright takedowns
On the US Semiquincentennial, historian Bret Devereaux reads the Declaration of Independence closely — a short but explosive 1776 text whose assertions of equality of rights and government by consent now seem obvious yet were deeply radical, and which 'remade the whole world.'
→ Reflective coda linking careful reading of foundational texts to broader perspective on current change

Questions addressed

✅ Is agentic development still speculative or has it become established practice?
Established practice — the 'whether' debate is over and the participants were shipping production, not slides.
⭕ Do architecture and design still matter when a powerful model can absorb poor structure?
Presented as an unresolved dispute between the 'Galaxy Brain' and 'DevEx = AgentEx' hypotheses; Fowler leans toward design still mattering but does not settle it.
⭕ How can design quality be measured for agent-driven work?
Token cost is proposed as a proxy, but the value of good architecture 'only shows over time' and is hard to measure short-term.
✅ Is renouncing AI the ethical choice?
No — there is no ethical gain from renouncing AI; engagement and responsible building do more for those harmed than abstention.
⭕ How should one ethically engage with AI in practice?
Left thin — Fowler concedes that neither Majors' suggestions nor his own retreat session produced anything striking.

Gaps

incompleteThe claim that fewer tokens indicates better architecture is asserted without any methodology to isolate architecture from confounds (model choice, prompt strategy, task variance, harness), so the proposed proxy is unvalidated.
Read next: Read Adam Tornhill's 'code for humans and machines' and look for empirical studies correlating architectural metrics with agent token cost.
closed byThe Economic Benefit of Refactoring — The gap calls for a methodology isolating architecture from confounds and for empirical studies correlating architectural metrics with agent token cost. This article runs exactly that kind of study: a fixed representative task executed by fresh, memoryless sub-agents at each of 15 refactoring steps on the same codebase and model, isolating code structure as the manipulated variable and reporting an 83% input-token drop. It does not fully resolve the confound question (n=1 codebase/task/model, no per-step variance), but is a direct, named first empirical answer to the gap's own reading direction ('look for empirical studies correlating architectural metrics with agent token cost').
illogicalThe post raises the 'Galaxy Brain vs DevEx = AgentEx' dispute and then leans toward the design-still-matters side without supplying evidence that adjudicates between the two hypotheses — the conclusion outruns the stated premises.
Read next: Seek controlled comparisons of agent performance and cost on well-structured vs. poorly-structured codebases.
closed byThe Economic Benefit of Refactoring — The gap asks for controlled comparisons of agent performance and cost on well-structured vs. poorly-structured codebases. This experiment is precisely that: the same codebase, the same task, and (as far as stated) the same model, with only the code's structure changing across refactoring stages — from a single 17,155-line file to 19 files with the largest under 3,700 lines — measured against the identical representative change at every stage.
uninformedAs an explicitly non-conclusion-forming fragments post, its picture of the 'state of practice' rests on a single retreat and a handful of blog posts rather than systematic adoption or outcome data.
Read next: Look for DORA-style empirical studies of agentic-development outcomes at scale.
incompleteThe ethics fragment argues for engagement over purity but concedes its concrete suggestions 'aren't striking', leaving the actionable ethical program underspecified.
Read next: Read Charity Majors' full 'Make AI Boring Again' essay and other work on responsible AI-in-practice for concrete engagement tactics.

Fragments: July 6

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents