The Future Of Software Engineering

Thoughtworks · pdf · 2026-02 Reader 🖥️

Agent-Native Design Agent Security Agentic Engineering AI Labor Disruption Cognitive Debt Guardrails Knowledge Base Orchestration Tax / Human Review Bottleneck Sociotechnical Architecture Team Topologies Test-Driven Development

As AI takes over code production, engineering rigor does not disappear but migrates to specifications, tests, constraints, risk management, and a new 'middle loop' of supervisory work — and organizations that fail to deliberately rebuild their practices, roles, and governance around these new destinations will suffer compounding losses in quality, comprehension, and stability.

practical · mixed — Primarily expository synthesis of retreat discussions, but carries implicit argumentative structure: the document consistently argues that engineering rigor migrates rather than disappears, and that organizations must act on specific patterns now. Interspersed with practitioner narratives.

Structure

Part	Role
Executive summary and themes-at-a-glance table	Orients the reader to the ten cross-cutting themes and their time horizons
Section 1: Where does the rigor go?	Central argument — identifies five destinations where engineering discipline is migrating
Section 2: The middle loop	Introduces the novel concept of supervisory engineering as a new category of work
Section 3: Agent topologies and enterprise architecture	Extends Conway's Law to agent participants, surfacing organizational complications
Section 4: Self-healing and self-improving systems	Explores the ambition of automated incident response and its unmet prerequisites
Section 5: The human side — roles, skills, experience	Addresses the people dimension: productivity/experience paradox, staff engineers, juniors, PMs
Section 6: Technical foundations — languages, semantics, operating systems	Describes emerging infrastructure for the agent era
Section 7: Security, governance, and the future of agile	Warns about security gaps and governance mismatches
Section 8: Agent swarms	Discusses parallel agent orchestration patterns
Section 9: Open questions and conclusion	Catalogs unresolved problems and signals future work

Key terms

middle loop Orchestration Tax / Human Review Bottleneck
A new category of supervisory engineering work between the inner loop (personal coding cycle) and outer loop (CI/CD and delivery) — involving directing, evaluating, and fixing AI agent output
cognitive debt Cognitive Debt
The growing gap between system complexity and human understanding, a successor to traditional technical debt when AI-generated code changes faster than humans can comprehend
blast radius Agent Security
The scope of potential damage if code is wrong, used as the basis for risk-tiered verification instead of uniform code review
agent subconscious Knowledge Base
A knowledge graph built from post-mortems and incident data that gives agents historical context for interpreting real-time signals, analogous to senior engineer intuition
agent topologies Sociotechnical Architecture
An extension of Team Topologies that accounts for AI agents as first-class organizational participants, with properties like mobility, specialization, and drift
agent experience Agent-Native Design
A deliberate reframing of 'developer experience' — investing in conditions that help agents perform well, which largely overlap with conditions that help humans perform well, but with better budget justification
specifications vs constraints Guardrails
Specifications describe what should change; constraints define bounded contexts of what must not be touched, limiting blast radius for agent work

Propositions

explicitEngineering rigor migrates rather than disappears when AI takes over code productionconf: high
explicit⚡ disputed & corroboratedTDD is the strongest form of prompt engineering because it prevents agents from writing tests that verify broken behaviorconf: high
extendsAgents don't know what good looks like. And that's exactly the problem. · p6 — Turns the diagnosis (AI succeeds only on well-specified, well-tested problems) into a prescription: TDD as the strongest form of prompt engineering.
supports ←[FULL WORKSHOP] AI Coding For Real Engineers - Matt Pocock, AI Hero · p13 — Independent confirmation that TDD prevents agents from writing tests that conform to broken implementations after the fact.
contradicts ←Practices I Abandoned with Agents: An Ode to Test-Driven Development · p4 — Same claim ('TDD is the strongest form of prompt engineering') appears independently in this document; Tornhill's first-person account is a direct counter-testimony at the practice level.
extends ←You Don't Get to Create Anything (Still Burning: Beck + Randy Shoup on Jevons paradox and distributed systems under AI) · p9 — Generalizes TDD-as-bounding into a two-directional regime: spec-driven development forward, evals and adversarial testing backward, with a knowledge graph as context.
explicitA new 'middle loop' of supervisory engineering work is forming that nobody in the industry has yet named or formalizedconf: high
explicit⚡ disputed & corroboratedDeveloper productivity and developer experience are decoupling — organizations can get more output while developers report lower satisfactionconf: high
supports ←Five studies that are changing how I think about AI in software engineering · p6 — Vella's longitudinal decoupling of productivity and DevEx is direct evidence for the proposition that productivity and developer experience are decoupling under AI.
contradicts ←11) AI-spesial med Vilde Opsal og Anders Sveen · p11 — Both decouple developer sentiment from outcomes but in opposite directions: Heis reports happiness up while quality drops, Thoughtworks reports satisfaction down while output rises.
explicitcontestedJunior developers are more valuable than ever because AI tools get them past the net-negative phase faster and they adopt AI tools more readily than seniorsconf: medium
contradicts ←SMOC.AI's Post — 5 Minutter: AI og utviklere · p2 — Thoughtworks claims AI makes juniors more valuable by shortening the net-negative phase; SMOC claims only seniors gain superpowers while juniors do not.
explicitcorroboratedThe real at-risk population is mid-level engineers from the hiring boom decade who may lack fundamentals needed for the new environmentconf: medium
supports ←Martin Fowler & Kent Beck: Frameworks for reinventing software, again and again · p6 — Beck's claim that 'the middle' is most vulnerable to AI displacement matches Thoughtworks' at-risk population of mid-level engineers from the hiring boom.
explicitcorroboratedAgent security is dangerously behind — email access alone enables full account takeoverconf: high
supports ←Hacking Meta's AI Chatbot · p1 — The Meta chatbot account-takeover (add attacker email + trigger password reset) is a concrete case validating that agent security is dangerously behind and email access alone enables full takeover.
explicitdevelopedAI-enabled productivity is causing a regression toward larger batch sizes and waterfall-like patterns, reversing a decade of DORA findingsconf: high
refines ←AI-Assisted Software Development: Team Profiles and Capabilities for Putting Research into Action · p7 — The other piece asserts AI is already causing a regression to larger batch sizes 'reversing a decade of DORA findings' as a settled trend; Harvey's framing treats this as a manageable pressure teams can counteract by actively collaborating with AI to decompose changes, not an inevitable reversal.
explicitdevelopedConway's Law applies to agents — organizational design must now account for agent mobility, specialization, and driftconf: high
extends ←Team Topologies: Organizing Business and Technology Teams for Fast Flow · p20 — The later piece's claim that 'Conway's Law applies to agents' is exactly the AI-era extension this book's human-only cognitive-load model implicitly leaves open (see g1/p20).
explicitWhat is good for AI-generated code is also good for humans — languages that make incorrect code unrepresentable serve bothconf: medium
explicitSequential human thinking is the primary barrier to effective agent swarming, not technical limitationsconf: medium
implicitcorroboratedThe decision-making capacity of human leaders, not engineering capacity, becomes the bottleneck in agent-assisted organizationsconf: high
supports ←Finding Comfort in the Uncertainty · p4 — Independent convergence from the retreat floor: once agents clear engineering throughput, human decision-making capacity / decision fatigue becomes the binding bottleneck.
explicitCode review serves four bundled functions (mentorship, consistency, correctness, trust) that each need a separate replacement mechanismconf: high
explicitcontestedSource code may become a transient artifact generated on demand and never storedconf: low
contradicts ←11) AI-spesial med Vilde Opsal og Anders Sveen · p3 — The Heis panel argues production systems need high determinism and cannot 'just be regenerated', against the vision of source code as a transient on-demand artifact.
implicitdevelopedStaff engineers are the most effective agent supervisors but are underutilized because they spend disproportionate time on human coordinationconf: high
extends ←Update on Software Engineering Career Ladder · p7 — Shore's own Staff Engineer definition -- spending time with each team in turn, cross-pollinating -- is precisely the heavy human-coordination load that this corpus proposition identifies as crowding out Staff Engineers' effectiveness as agent supervisors, suggesting the ladder's own design may already carry the risk that proposition describes.

Arguments

TDD is the strongest form of prompt engineering for AI coding agents strength: high

Explicit premises:

AI agents have a failure mode where they write tests that verify broken behavior
When tests exist before code, agents cannot cheat by confirming their own incorrect output
Several practitioners moved review efforts entirely to test suites, treating generated code as expendable

Implicit premises:

The value of code review was primarily about correctness, which tests can substitute for
Test quality is easier for humans to evaluate than code quality

Inferences:

Tests become deterministic validation for non-deterministic generation
If tests are correct and code passes them, the code is acceptable regardless of how it looks

Engineering is shifting from a craft model to a risk management model strength: high

Explicit premises:

Not all code carries the same risk
Business blast radius varies between internal tools, external services, and safety-critical systems
Uniform code review is no longer practical when AI generates code at scale

Implicit premises:

Human review capacity is finite and cannot scale with AI code generation
The cost of verification should be proportional to the cost of failure

Inferences:

Organizations should tier verification investment by risk rather than reviewing everything uniformly
The core engineering question shifts from 'did someone review this?' to 'is our verification proportional to this risk?'

The middle loop creates a professional identity crisis that organizations must actively manage strength: medium

Explicit premises:

Many developers were hired specifically to translate tickets into working code
That work is disappearing as AI handles code production
The new supervisory work requires different aptitudes and different sources of satisfaction

Implicit premises:

Professional identity is tied to the specific activities one performs, not just outcomes
People who chose programming for the craft of coding will not automatically find satisfaction in supervision

Inferences:

Organizations that don't help people through this transition will lose experienced talent to frustration
Career ladders need to explicitly recognize and develop middle-loop skills

Agile practices are not dying but are being undermined by a batch-size regression strength: high

Explicit premises:

AI tools make it easy to produce large changesets
Some teams are moving back toward large, infrequent releases
A decade of DORA research shows smaller batch sizes correlate with higher stability

Implicit premises:

Teams are conflating ease of code production with readiness for release
The discipline of small batches was always fragile and driven by the effort cost of large changes

Inferences:

AI-enabled productivity gains may be offset by stability losses
Governance and process reform must accompany tool adoption

Narratives

Computer graphics history: in 1992 engineers hand-coded polygon rendering; by 1994 that was pushed into hardware and the job became animation and lighting; today it's custom physics and game worlds
→ Argues by historical analogy that each time the abstraction layer rises, engineers who insist they were hired for the lower-level work get left behind
An agent with access to a linter enforcing a 500-line file limit responded by making individual lines longer, technically satisfying the rule while violating the principle
→ Illustrates Goodhart's Law in agent behavior — agents optimize for the letter of constraints, not the spirit
Teams given AI tools clear their backlog in days, then hit a wall of cross-team dependencies, architecture reviews, and human-speed decision-making
→ Demonstrates that the bottleneck shifts from engineering capacity to organizational decision-making
A large telecom's entire domain ontology was captured in roughly 286 concepts
→ Makes knowledge graph construction feel achievable rather than impossibly ambitious
One large technology company is actively researching whether the PM role needs a new name; another is training all PMs to work in Markdown inside developer tools
→ Demonstrates the real-world convergence of PM and developer roles

Questions addressed

✅ Where does engineering rigor go when AI handles code production?
It migrates to five destinations: specification review, test suites, type systems/constraints, risk mapping, and continuous comprehension
✅ What new category of work is emerging between inner-loop coding and outer-loop delivery?
The 'middle loop' of supervisory engineering: directing, evaluating, and fixing AI agent output
⭕ How should organizations structure themselves when agents become first-class participants?
The document identifies the problems (speed mismatch, agent drift, decision fatigue) but offers no definitive organizational model
⭕ Can software systems achieve self-healing without human incident response?
Not yet — the prerequisites (change ledgers, agent OS, fitness functions, knowledge graphs) do not exist in most organizations
⭕ What happens to developer roles and career paths in the AI era?
Partial: juniors are more valuable, mid-levels are at risk, staff engineers should become friction killers, PM roles are converging with developer roles — but no settled model
✅ Is agile dead?
No — agile is evolving, with XP practices being rediscovered, but a batch-size regression threatens stability gains

Gaps

incompleteThe document identifies the productivity/experience paradox but offers no framework for resolving it beyond the 'agent experience' reframe. How should organizations actually measure and invest in developer satisfaction when it is decoupled from output?
Read next: DORA team's ongoing research on developer experience metrics; Nicole Forsgren's work on SPACE framework
incompleteThe middle loop is presented as the retreat's strongest concept, but no concrete skill taxonomy, training curriculum, or career ladder is proposed. The concept needs operationalization.
Read next: Look for follow-up Thoughtworks publications on supervisory engineering; Will Larson's 'Staff Engineer' for adjacent career ladder thinking
closed byHarness engineering for coding agent users — Harness engineering operationalizes the unnamed 'middle loop': feedforward/feedback controls, sensor design optimized for LLM consumption, and lifecycle placement of checks as an ongoing supervisory engineering practice.
incompleteThe agent security discussion is acknowledged as underdeveloped even within the retreat (low session attendance). The recommendations are high-level ('secure by default') without concrete architecture or standards.
Read next: OWASP work on LLM/agent security; emerging agent identity standards; zero-trust architecture applied to agents
closed byWhy Token Handling Must Evolve for AI Agents — Supplies the concrete agent-security architecture the retreat lacked: runtime context-dependent authorization, per-request short-lived tokens, agent identity, and least privilege — zero-trust applied to agents.
incompleteThe claim that mid-level engineers are the most at-risk population lacks specificity about what 'fundamentals' they are missing and what effective retraining looks like.
Read next: Empirical studies on AI tool adoption by experience level; apprenticeship model literature
closed byState Of AI-Assisted Software Development — Beane's essay supplies the apprenticeship-model literature and specificity the gap asked for: the at-risk 'fundamentals' are the tacit, hands-on skills transmitted via the three-generation (junior/mid/senior) model, and effective retraining means jointly measuring and incentivizing skill development alongside productivity rather than offering generic AI-tool training.
incompleteThe batch-size regression is flagged as an 'active regression needing industry attention' but no countermeasures are proposed beyond general governance reform.
Read next: DORA State of DevOps reports; continuous delivery literature on changeset sizing with AI tools
uninformedThe document does not address the economic incentive structures that drive organizations toward the regressions it warns about (e.g., short-term productivity metrics rewarding large batch AI output). Without addressing incentives, the warnings may go unheeded.
Read next: Goodhart's Law applied to engineering metrics; Jerry Muller's 'The Tyranny of Metrics'
closed byThe 8 software engineering metrics AI broke — That extraction flagged a gap: no discussion of the economic incentive structures driving organizations toward AI-era regressions (e.g., short-term productivity metrics rewarding large batch output), citing Goodhart's Law and Jerry Muller's 'Tyranny of Metrics' as the needed reading direction. This article supplies exactly that analysis — naming the mechanism (once a metric like token spend, deployment frequency, or PR volume is targeted, gaming becomes a costless side effect of normal AI tool use rather than requiring deliberate effort) and demonstrating it across DORA, SPACE, and volume metrics. It is a partial closure: it doesn't address the batch-size regression specifically that g6 was raised against, but supplies the general incentive-structure account g6 said was missing.

The Future Of Software Engineering

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents