Finding Comfort in the Uncertainty

Annie Vella · article · 2026-07-07 source ↗ Reader 🖥️

Agent Governance (Internal)Agent-Native Design Agent Work Ledger AI as Amplifier AI Employee vs Tool Framing Code Was Never the Bottleneck Cognitive Debt Cognitive Load Enterprise AI Adoption & Readiness Human as Orchestrator of Agents Orchestration Tax / Human Review Bottleneck Platform Engineering Supervisory Programming

The most valuable takeaway from a gathering of leading software practitioners on AI's impact is that nobody has it figured out — and that this shared uncertainty, together with a convergent set of hard questions the field is only beginning to name, is itself the ground for progress and a source of comfort rather than alarm.

mixed · mixed — A first-person reflective synthesis of an invite-only retreat on the future of software development: narrative framing (the journey home, the room in Deer Valley) wrapped around an expository survey of eight convergent themes, closing with a normative call to action.

Structure

Part	Role
Opening frame: the flight home and Deer Valley	Shows the retreat's questions surfacing unprompted between strangers on a plane, then situates the event in the mountains where the Agile Manifesto was written 25 years ago — signalling that the ground has shifted enough to warrant a comparable conversation
Nobody Has This Figured Out	States the author's single most important realisation and reframes the pervasive uncertainty as comforting and the shared question-set as the retreat's most valuable outcome
The Increasing Cognitive Load to Be Aware Of	Surfaces the hidden human cost — the addictive spin-up of parallel agents and the resulting review overload burning out the most senior, enthusiastic engineers
What I Brought to the Table	Positions the author's own Masters-research contributions (supervisory engineering work; possible divergence of productivity and developer experience)
Eight Themes That Kept Emerging	The expository core — catalogues the questions that independently converged across 30+ sessions (moved bottleneck, artifact primacy, trust/care/debt, platform enablement, agent taxonomy, organisational readiness, redefined human role, the ledger)
Why It Mattered	Normative close — the challenge to describe the new problems engineers get to solve rather than offer platitudes, and the commitment to keep the discourse open

Key terms

supervisory engineering work Supervisory Programming
A new category of engineering work — supervising, orchestrating, and evaluating AI output — emerging as implementation is delegated to agents; the author's coinage from her Masters research
the ledger (work ledger)
A complete, verifiable record of everything agents do — every step and change, with the ability to roll back — variously framed as a work ledger, an audit trail, or a unified log
cognitive debt Cognitive Debt
Knowledge losing its stickiness because AI unlocks so much volume that we skim everything — an extension of technical debt into the mind
trust debt
The gap between what we produce and what we deeply understand about what we've produced
active knowledge decay
Skills atrophying when you stop doing the work hands-on
Agent Experience Agent-Native Design
Investment in the conditions that make agents effective (clear standards, good documentation, well-structured code) — the same conditions that made developers effective, hence possibly 'the new DevEx'
middle loop Orchestration Tax / Human Review Bottleneck
An emerging layer of work — supervising, orchestrating, evaluating — sitting between the increasingly automated inner loop of coding and the outer loop of delivery
pit of success
Platform defaults that make the safe path also the fast, easy path so that speed and safety are not in tension

Propositions

explicitEven the sharpest minds in software have not figured out how to adopt AI at scale; there is more uncertainty than certainty about how to use AI well, its productivity impact, and how roles are shifting.conf: high
explicitA shared understanding of the right questions to ask about AI in software development is more valuable than premature answers.conf: high
explicitThe engineers going hardest on AI are being exhausted by it: spinning up multiple parallel coding agents is addictive, but their outputs return at once as pages of review, and because those hit hardest are the most senior and enthusiastic, the warning signs are easy to miss.conf: high
explicitIn some organisations the binding constraint is no longer engineering capacity but downstream absorption — slow decisions, organisational dependencies, and limited market appetite ('humanity is not ready for this much software') — so decision fatigue, not engineering throughput, is becoming the real bottleneck.conf: high
supportsThe Future Of Software Engineering · p12 — Independent convergence from the retreat floor: once agents clear engineering throughput, human decision-making capacity / decision fatigue becomes the binding bottleneck.
explicitIf agents can regenerate code from a specification, source code may no longer be the durable artifact — the domain model, the test suite, or the intent may be what endures, and code becomes a disposable 'projection' of intended behaviour.conf: medium
explicitAI-driven volume produces new forms of debt beyond technical debt — cognitive debt, trust debt, and active knowledge decay — which are problems of identity, meaning, and accountability, not just process.conf: high
explicitPlatform engineering that provides safe defaults (the 'pit of success') and embedded guardrails gives teams a fast-but-safe path, dissolving the assumed tension between speed and safety.conf: high
explicitThe practices that make AI agents effective (clear standards, good documentation, well-structured code) are the same ones that made human developers effective, so investing in 'Agent Experience' may be the new Developer Experience and humans may quietly benefit too.conf: medium
explicitThe industry has no settled mental model for agents (tool/API vs. team member/human vs. non-deterministic 'chemical vat'), and this ambiguity is slowing coherent governance and investment decisions.conf: high
explicitAI amplifies existing organisational conditions rather than uniformly improving them — mature organisations with strong teams, platforms and governance get faster, while those without foundations get messier — so the first year of adoption should focus on preparing the system rather than expecting immediate productivity gains.conf: high
supportsPresentation: AI-First Software Delivery: Balancing Innovation with Proven Practices · p5 — Restates the amplification thesis as a convergent retreat finding: AI magnifies existing conditions — strong foundations get faster, weak ones get messier.
explicitcorroboratedA new category of 'supervisory engineering' work — supervising, orchestrating, and evaluating AI output in a 'middle loop' between the automated inner loop of coding and the outer loop of delivery — is emerging, but its skills, career pathways, and even its dignity have not yet been designed.conf: high
supports ←What's an Orchestrator—and Why Does Software Need One? · p3 — O'Brien's 'orchestrating the systems that operate and develop software' is a concrete first-person instance of the 'supervisory engineering' middle-loop category identified there.
explicitMultiple independent groups converged on the need for a complete, verifiable ledger of everything agents do — every step and change, with the ability to roll back — as a precondition for trusting agents with real work.conf: high
explicitThe profession's obligation now is to get better at describing the new problems engineers get to solve — not to reassure them that everything will be fine — because writing code is no longer the main problem.conf: high
implicitThe field needs new shared language, principles, and ways of thinking, but these cannot yet be authored — they can only emerge from continued open discourse.conf: medium

Arguments

The shared uncertainty of the retreat is a source of comfort and progress rather than alarm. strength: medium

Explicit premises:

Nobody — not even the sharpest minds present — has AI-in-software figured out; there is more uncertainty than certainty.
The group nonetheless left with a shared understanding of the sorts of questions it should be asking.
That shared question-set 'might be the most valuable outcome of all'.

Implicit premises:

Naming the right questions is a precondition to answering them.
Honest collective inquiry is preferable to isolated false certainty.

Inferences:

Therefore discovering that nobody has the answers, but everyone shares the questions, is itself the valuable and reassuring result.

Organisations should spend the first year of AI adoption preparing their foundations rather than chasing immediate productivity gains. strength: high

Explicit premises:

AI amplifies existing conditions: organisations with strong teams, platforms and governance got faster; those without got messier.
Microservices taught the same lesson — without a certain level of organisational maturity the technology makes things worse, not better.
Several groups independently converged on prioritising platform foundations, governance, and AI-fluency uplift first.

Implicit premises:

Organisational readiness is a prerequisite for value, not a byproduct of adoption.
Expecting broad productivity gains before the foundations exist causes harm rather than benefit.

Inferences:

Therefore foundation-building, not productivity, is the correct first-year goal.

The profession must deliberately design the emerging human role rather than stumble into it. strength: medium

Explicit premises:

Everyone agrees the software engineer's role is shifting, but nobody has a design for what it shifts to.
A supervisory 'middle loop' category of work is emerging, yet its skills, pathways, and dignity are undesigned.
Many engineers hired to write code are now being told to outsource that to AI and deserve more than platitudes.

Implicit premises:

Undesigned transitions leave people without pathways or dignity.
Those with platforms and voices bear a responsibility to shape the transition.

Inferences:

Therefore leaders should do the hard work of defining the new problems and pathways rather than offering reassurance.

Narratives

On the final flight home the author is wedged between a retired industrial chemist and a theoretical physicist reading a C++ book; the physicist thinks AI-generated code is 'pretty crap' yet has noticed engineers around him losing their jobs.
→ Opens the piece by showing the retreat's questions surfacing unprompted among strangers — the concerns are universal, not an insider preoccupation
About 40 practitioners, researchers and technical leaders gather in Deer Valley, Utah, in the same mountains where the Agile Manifesto was written 25 years ago, with some original signatories present — an unconference with no presentations and no handed-down agenda.
→ Establishes the stakes and legitimacy of the moment while explicitly disclaiming any attempt to reproduce the Agile Manifesto
Between sessions, several people going hardest on AI quietly admit that it is exhausting them, with several independently using the word 'addictive' about spinning up parallel agents.
→ Humanises the cognitive-load thesis; the private honesty is presented as the emotional core of the retreat
At the close, when the mic is offered for takeaways, the author challenges the group — all with platforms and voices — to help engineers by getting better at describing the new problems they get to solve, not by reassuring them that everything will be fine.
→ Delivers the normative call to action that carries the article's thesis into an obligation

Questions addressed

✅ Has anyone figured out how to adopt AI in software development at scale?
No — even the sharpest minds in the room had not; there is more uncertainty than certainty, and everyone is working it out as they go.
✅ Once agents can burn through backlogs, what becomes the real bottleneck?
Downstream human and organisational absorption — slow decisions, dependencies, market appetite — with decision fatigue replacing engineering throughput as the constraint.
⭕ If code can be regenerated from a specification, what is the durable artifact?
Left open — candidates offered include the domain model, the test suite, and the intent, but no answer is settled.
⭕ How should the industry conceptualise agents — as tools, team members, or something else?
Unresolved — no settled mental model exists (API vs. human vs. non-deterministic 'chemical vat'), and the ambiguity is slowing governance and investment.
⭕ What does the software engineer's role shift to?
Undesigned — a supervisory 'middle loop' is emerging, but its skills, pathways, and dignity have not been shaped; the field is stumbling into it.

Gaps

incompleteThe 'supervisory engineering' role is reported as emerging but is not designed: the piece names the category without specifying its actual skills, career pathways, evaluation criteria, or how its dignity would be established.
Read next: Read Annie Vella's Masters research on supervisory engineering work and look for concrete role-design or career-ladder frameworks for the human-as-orchestrator layer.
incompleteThe 'ledger' is asserted from convergence but its mechanism, cost, and governance are unspecified — what a complete, verifiable record of agent actions actually records, how rollback is guaranteed, and how it is enforced across infrastructure, software, and network layers.
Read next: Read on agent audit trails, decision provenance, and policy-as-code approaches to enforcing verifiable agent action logs.
illogicalThe reframing of pervasive uncertainty as 'comforting' may be a non-sequitur for the wider profession: the comfort felt by invited insiders sharing questions does not obviously transfer to the engineers, mentioned in the same piece, who are losing jobs or being told to outsource their core work now.
Read next: Read on AI labor disruption and the lived experience of displaced or de-skilled engineers to test whether shared inquiry actually reassures those most exposed.
closed byLenny Rachitsky's Post — That gap suspects the 'comfort in uncertainty' felt by invited insiders does not transfer to the exhausted wider profession. This survey quantifies the felt distribution: ~59% (Conflicted 35% + Disoriented 12% + Resentful 12%) sit in overwhelmed/tired/disoriented/resentful states, and the Disoriented archetype in particular (VP Product: 'so uncertain... crowds out productivity') is a direct empirical counterexample — showing uncertainty is a source of drag and distress for most, not comfort.
incompleteThe claim that productivity and developer experience — long assumed to move together — might be starting to diverge is raised as one of the author's own topics but is offered without evidence, measure, or mechanism.
Read next: Read on measuring AI impact and developer-experience metrics to see whether the divergence is empirically observable.
closed byFive studies that are changing how I think about AI in software engineering — That gap flags that the productivity/developer-experience divergence was raised without evidence, measure, or mechanism. Vella's six-month longitudinal study of 95 engineers supplies the evidence and measure: the worse-DevEx share nearly doubled (14% to 27%) while perceived productivity stayed high (84%), and productivity change-scores did not correlate with DevEx change-scores. (It supplies measurement, not a deep causal mechanism.)

Finding Comfort in the Uncertainty

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents