Rewiring software delivery for the agentic era

Moon / Walsh / Di Leo — Rewiring software delivery for the agentic era (McKinsey Technology, 2026-05-28)

The way agentic AI is being used in software development is a harbinger for broader changes in the delivery model.

(Article subhead, McKinsey & Company / McKinsey Technology.)

A ~7-page article from McKinsey Technology, published 28 May 2026. Lead authors: Jared Moon (Senior Partner, London office), Rory Walsh (Partner, Dublin), Vito Di Leo (Partner, Zurich), with Adam Thelwall (Associate Partner). Acknowledged contributors: Aishik Dhar, Amray Schwabe, Benjamin Schloesing, Nikolaus Müller. Edited by Barr Seitz (Editorial Director, New York).

This is the wiki’s second McKinsey-published source on agentic-era enterprise transformation — joining Sternfels (Feb 2026) at the firm-level-reinvention altitude with a tech-partner-operational-altitude specification of how the reinvention happens inside one work domain (software delivery). Also part of the McKinsey & Company post-controversy intellectual recovery — three McKinsey-internal vantages (Sternfels managing-partner / MGI macroeconomic / McKinsey-Digital tech-partner) now anchor the firm’s AI-rewires-the-operating-model doctrine inside the wiki.

TL;DR

Six substantive contributions:

1. The 9 a.m. vignette and the 24-hour-sprint model — the article’s organising operational frame. “At 9:00 a.m., a product owner logs in to review overnight progress on a solution her team is working on. She sees that a feature has moved from structured requirements to tested code. Edge cases are flagged. She notes that architecture dependencies have been validated. A concise summary outlines trade-offs and open decisions. No one worked late. AI agents did.” The operational structure (Exhibit 1):

Shift	Hours	Activities
Night shift (led by a factory of agents)	16 hours	Requirements: create business requirements for the features requested by humans. Architecture: check if architecture is in place; set up structure; create design. Build and test: build and test a first version; write a report with outcomes and recommendations.
Day shift (humans supported by agents)	8 hours	Sprint review/demo: review agent output vs expectations and acceptance criteria. Spec-and-code working session: live pair review of critical code paths + AI traceability + cross-functional sync (legal/compliance inputs, design feedback). Offline system optimization: refactor weak code flagged in morning session; refine guardrails and quality standards for context, skills, prompts, and workflows; rerun the factory; design improvements for next sprint. Sprint planning: refine inputs and instructions for the agentic factory; align with stakeholders on priorities for the upcoming night shift.

The framing of the human role: “Increasingly, their role is less about producing artifacts and more about supervising and improving the system that produces them.” The wiki’s clearest single-sentence McKinsey-altitude declaration of the human role shifts from producing to supervising the system that produces claim — convergent with [[2026-05-21-allen-aws-london-exec-forum-agentic-team-structures|Allen’s USE / COMPOSE / BUILD framing]] and with [[2026-05-27-koomen-yc-lightcone-inside-yc-ai-playbook|Koomen’s technology becomes leading, smaller group of people directs it]] from the day before.

Four practical foundations for the 24-hour model: (1) business must have clear vision of what needs to be built (product road map / standard to build from); (2) underlying technology environment must be standard and consistent (common frameworks, modular architectures); (3) path from requirements to code must follow a standard structure; (4) same core stakeholders need to stay engaged across the value stream. Main takeaway: “Continuous 24-hour delivery is achievable but only with architectural discipline and standardized workflows so agents can operate reliably at scale.”

2. The extend automation to eliminate human handoffs thesis with the 30%/70% CI/CD cost decomposition. “Traditional continuous integration and continuous delivery (CI/CD) automation focuses largely on testing and deployment. While those costs vary, our experience is that they can be as much as 30 percent of total technology spend. The majority of effort, concentrated in requirements through coding, remains manual and interpretation heavy. This is where friction accumulates and value plateaus.” The agentic model removes this friction by structuring artifacts for machine-to-machine handoffs — functional descriptions, nonfunctional requirements, guardrails, sequence diagrams, repositories codified in standardized machine-readable formats. “The pipeline can then run end to end in hours, with humans intervening only at defined review gates rather than acting as intermediaries.” Main takeaway: “Scaling AI requires applying engineering practices to the development system itself, making the process repeatable and automating handoffs.” The McKinsey-altitude 30%/70% deployment-vs-requirements-through-coding cost decomposition is the wiki’s clearest quantitative framing of where the agentic SDLC actually attacks cost — the 30% (CI/CD) is already automated; the 70% (the interpretation-heavy upstream pipeline) is where the agentic-factory wins live.

Exhibit 2 (the agent-squad pipeline diagram) decomposes the path: Requirements (project brief / business requirements / functional descriptions / external interface specs → high-level process flow / enriched requirements / refined business requirements) → Design (future-state architecture / technical guardrails / nonfunctional requirements / code repositories / rules catalog → summarized standards and design / technical user stories / sequenced diagrams and specifications) → Code/test (development guidelines / YAML rules framework → business validation services pull request / message transformation services pull request / test report) → Deploy (traditional CI/CD automation). Sources cited: “Software development cost: Complete 2026 budget guide,” Boundary AI, Apr 2, 2026; McKinsey analysis.

3. The knowledge graphs as AI memory layer infrastructure prescription — the wiki’s clearest consulting-altitude framing of knowledge as production infrastructure. “To produce accurate results, agent factories need organizational context and memory. Top businesses are building knowledge graphs that function as an AI memory layer across the software development life cycle (SDLC) for each domain. These graphs connect elements that agents need to make sense of, such as customer feedback, architecture decisions, design documents, tickets, GitHub activity, incident reports, and summarized compliance rules. The result is a semantically linked system (that is, a way for agents to understand what the data means so they can better perform their tasks).”

The wiki’s clearest named role and worked example for the librarian agent: “Questions that once required weeks of interviews with multiple subject matter experts (SMEs) can be answered in minutes by a ‘librarian’ agent drawing from structured institutional memory. Every decision becomes traceable. If a stakeholder asks why a feature was deprioritized, the answer can be linked directly to its source, such as customer survey data or usage analytics. Implicit tribal knowledge becomes explicit and explainable, reducing ramp-up time for new team members and strengthening governance.”

Critical anti-grand-ontology design principle: “Importantly, this should not begin with a grand, top-down ontology effort. The graph should evolve organically around priority domains and live programs, compounding value over time. As it scales, knowledge becomes production infrastructure, rather than static documentation, and a durable source of competitive advantage.” Main takeaway: “Structured, connected knowledge is the foundation of agent autonomy. Treat your knowledge architecture as strategic infrastructure. Knowledge graphs are the critical unlock to enable velocity and agent autonomy.”

Exhibit 2 sample sources for the knowledge graph: SharePoint (previous-survey results — drop-off drivers), Observability (recent customer usage data and drop-off points), Jira (features / user stories already planned), GitHub (existing services supporting onboarding or related capabilities), SMEs (implicit knowledge on the journey, not formally documented).

4. The team-size quantification anchor — 100 → 60 FTEs / 200 → 100 person-years / 10 teams of 8-12 FTEs → 16 teams of 3-4 FTEs. Exhibit 3:

Dimension	Current delivery model	Agentic SDLC	Reduction
Full-time equivalents	~100 FTEs	~60 FTEs	~40%
Duration	200 person-years (24 months)	~100 person-years (18 months)	~50% total effort
Project team makeup	10 teams of 8-12 FTEs	16 teams of 3-4 FTEs	~60% average team size
Team composition	Product owner / Business analyst / Tech lead / 5 Software engineers / 2 Testers (= 10 roles)	Product owner / Tech lead / AI-enabled engineer (= 3 roles)	7 of 10 roles dissolved

Source: “Based on McKinsey experience and observation across multiple companies.” The wiki’s clearest McKinsey-consulting-aggregated team-size-reduction empirical anchor — convergent with AWS London (Project Mantle 76 days vs 18 months, embedded-pod model 3-5 engineers / pod, junior-hiring-crisis), AI Dev 26 SF (PM-bottleneck cascading-bottlenecks, small-team-of-generalists), and MGI’s 7-archetype + Skill-Change-Index framework. Headline article-grab claims: “multiple companies are already seeing it deliver threefold to fivefold improvements in productivity, with a 60 percent reduction in team size.”

5. The outer-loop roles integration prescription — risk / legal / testing / procurement baked into the agentic-development effort by design, not as end-of-process gatekeepers. “Second is ensuring the ‘outer loop’ roles — support and compliance people in risk, legal, testing, and procurement — are part of the agentic development effort. A faster SDLC doesn’t translate into faster progress if this doesn’t happen. Agents and automation (for example, through policy as code) can help to ensure these controls don’t become bottlenecks, while improving quality, consistency, completeness, and traceability. These controls should be baked in by design, rather than becoming a gatekeeper at the end of the process.” The wiki’s clearest McKinsey-altitude policy-as-code prescription as a named primitive for compliance integration. Convergent with Brooklyn Solutions’s regulated-customer harness composition (AWS Bedrock + Bedrock Guardrails + AgentCore, human-starts-human-ends accountability) but at consulting-altitude abstraction rather than vendor-customer worked-example altitude. Plausible single-source-deferred concept-page candidate: policy-as-code as outer-loop integration primitive.

6. The capacity-reinvestment-not-cost-cut normative framing — productivity gains must translate into structural portfolio changes. “Productivity gains can be translated into structural portfolio changes. Resize teams and consciously redeploy capacity to capture full value.” Three priorities: (1) reskill people — software engineers need judgment, code review, and supervisory skills to manage agents they work with; roles shift away from manual coordination and testing toward architecture coherence, domain modeling, and AI supervision; (2) outer-loop integration (item 5 above); (3) redesign how capacity is allocated so productivity improvements translate into new value: “Freed capacity is often reinvested to accelerate road maps, modernize platforms, or launch new products.” The McKinsey-consulting-altitude avoid-the-micro-productivity-trap normative argument — convergent with micro-productivity-trap’s framing and with the wiki’s MGI Yee workflow-not-task imperative.

Closing thesis: human roles concentrate in architecture, product judgment, and system design

“As agents take on execution at scale and produce code that is robust and consistently secure, human roles will concentrate in architecture, product judgment, and system design, making institutional knowledge and technical coherence decisive differentiators. Organizations that begin building these capabilities as part of a broader effort to rewire their operating model will not just move faster; they will redefine how software creates value.”

This is the wiki’s clearest McKinsey-consulting-altitude statement of the durable-skills-concentrate-at-the-top-of-the-K-curve claim — convergent with Elangovan’s K-shaped future of software engineering (top arm: systems-level thinking, judgment and taste, intuition, problem framing, harness setup) and with [[2026-05-21-sinclair-ivers-benitez-sei-cmu-ai-native-software-engineering|SEI/CMU’s coder-vs-software-engineer role-shift]] thesis. McKinsey at the consulting-altitude lands the same observation: human roles concentrate at architecture + product judgment + system design + institutional knowledge + technical coherence.

The transformation strategy prescription: “Transformation should begin where impact is greatest. In most technology organizations, a small number of large programs account for the majority of total spend. Targeting these initiatives — whether legacy modernization, brownfield rebuilds, or new product launches — maximizes visible impact and accelerates learning.”

Why this matters in the corpus

Three sub-corpus roles for this source:

1. The wiki’s second McKinsey-published source on agentic-era enterprise transformation — extends Sternfels (Feb 2026)‘s firm-level-reinvention thesis with a same-firm 3.5-month-later operational specification of how that reinvention happens inside one work domain. The McKinsey-internal cluster now spans three vantages (Sternfels managing-partner / MGI macroeconomic / McKinsey-Digital tech-partner).

2. The wiki’s clearest team-size-reduction quantification anchor at McKinsey-consulting-aggregated-experience altitude. The 100 → 60 FTEs / 200 → 100 person-years / 10 teams of 8-12 → 16 teams of 3-4 / 10-role-pod → 3-role-pod data is the most-citable McKinsey-altitude numerical anchor for the team-structure-redesign cluster (Allen / AWS; Ng / AI Dev 26 SF; Tan & Hu / Stanford CS153; Sinclair / SEI/CMU; Brynjolfsson / Canaries). Useful as quantitative anchor for ai-employment-effects and enterprise-ai-adoption discussions.

3. The wiki’s clearest knowledge-as-production-infrastructure consulting-altitude framing, including the named librarian agent role + the anti-grand-ontology design principle (the graph should evolve organically around priority domains, not begin with a top-down ontology effort) + the implicit-tribal-knowledge-becomes-explicit-and-explainable governance claim. The wiki’s existing knowledge-architectures-for-llm-agents synthesis (3 sources) and is-rag-dead synthesis (7 sources) both gain a McKinsey-consulting-altitude operational anchor.

The W&W tagging (9 cells — same breadth as De Ondernemer from the same day, comparable to JetBrains (10 cells), AWS London (10 cells), and YC Lightcone (8 cells)) reflects this source’s reach across the W&W process model — touching three of five buckets substantively. The presence of both strategic-renewal/business-model (“redefine how software creates value”) and strategic-renewal/organizational-culture (“role shifts from producing artifacts to supervising the system that produces them”) cells reflects this source’s reach into the strategic-renewal outcome layer.

Linked entities and concepts

Entities (already promoted, source_count bumped):

• McKinsey & Company — publisher; source_count 8 → 9.

Dangling first-mentions (single-source, deferred per §Lifecycle author-entity promotion):

• Jared Moon — Senior Partner, McKinsey London office. Lead author.
• Rory Walsh — Partner, McKinsey Dublin office.
• Vito Di Leo — Partner, McKinsey Zurich office.
• Adam Thelwall — Associate Partner.
• Aishik Dhar — acknowledged contributor.
• Amray Schwabe — acknowledged contributor.
• Benjamin Schloesing — acknowledged contributor.
• Nikolaus Müller — acknowledged contributor.
• Barr Seitz — Editorial Director, McKinsey New York office (editor).
• Boundary AI — cited as 2026-04-02 source for the Software development cost: Complete 2026 budget guide reference.

Concepts (last_confirmed bumped; substantive new content possible on agentic-engineering, enterprise-ai-adoption, ai-employment-effects, agent-development-lifecycle):

• agentic-engineering — the 24-hour-sprint day-shift / night-shift model + factory of agents + humans-supervise-the-system-that-produces-artifacts role-shift is the McKinsey-consulting-altitude operational specification of agentic engineering as a discipline.
• agent-harness — factory of agents + machine-to-machine handoff artifacts (functional descriptions, nonfunctional requirements, guardrails, sequence diagrams, repositories codified in standardized machine-readable formats) is convergent with the wiki’s harness-as-substrate framing.
• agent-development-lifecycle — Exhibit 2’s Requirements → Design → Code/test → Deploy agent-squad pipeline is a substantive McKinsey-consulting-altitude SDLC-redesign worked example for the ADLC concept.
• enterprise-ai-adoption — “rewiring the operating model” + “transformation should begin where impact is greatest” + the three-priorities operationalisation (reskill / outer-loop integration / capacity redesign).
• ai-employment-effects — the 100 → 60 FTEs / ~60% team-size reduction / 10-role-pod → 3-role-pod quantification + “freed capacity is often reinvested to accelerate road maps, modernize platforms, or launch new products” + the reskilling-not-displacement normative framing.
• micro-productivity-trap — “Productivity gains can be translated into structural portfolio changes. Resize teams and consciously redeploy capacity to capture full value.” — McKinsey-altitude avoid-the-trap normative argument.
• automation-vs-augmentation — Moon-Walsh-Di-Leo sit on the agents-execute-humans-supervise augmentation side at the structural level (humans concentrate at architecture / product judgment / system design) while pursuing aggressive automation at the execution layer.
• durable-skills — “human roles will concentrate in architecture, product judgment, and system design, making institutional knowledge and technical coherence decisive differentiators.” The wiki’s clearest McKinsey-consulting-altitude statement of the top-of-K-curve durable-skills cluster.

Synthesis pages gaining a McKinsey-consulting-altitude operational anchor:

• knowledge-architectures-for-llm-agents — the knowledge graphs as AI memory layer + librarian agent + anti-grand-ontology design principle + knowledge as production infrastructure / durable source of competitive advantage framing is a substantive McKinsey-consulting-altitude operational anchor.
• is-rag-dead — the agents-need-organizational-context-and-memory claim + knowledge-graph-as-substrate is convergent with the synthesis’s RAG-alive-inside-agentic-harness conclusion at McKinsey-altitude.

Concepts the wiki may want to promote following this source (single-source-deferred):

• 24-hour sprint model — McKinsey’s distinctive operational frame; defer to second-source promotion.
• Factory of agents — McKinsey’s named primitive; defer.
• Librarian agent — McKinsey’s named role for knowledge-graph-querying agent; defer.
• Policy as code as outer-loop integration primitive — McKinsey’s prescription for risk / legal / testing / procurement baked in by design; defer.
• Knowledge as production infrastructure (as durable competitive advantage) — McKinsey’s strongest single-line strategic-renewal claim about knowledge architecture; defer to second-source.

Source

• Raw PDF (7 pp): article file — browser-print save from mckinsey.com.
• Public URL: mckinsey.com/capabilities/mckinsey-digital/our-insights/rewiring-software-delivery-for-the-agentic-era
• Publisher: McKinsey & Company / McKinsey Technology.
• Authors: Jared Moon (Senior Partner, London) + Rory Walsh (Partner, Dublin) + Vito Di Leo (Partner, Zurich) + Adam Thelwall (Associate Partner).
• Editor: Barr Seitz (Editorial Director, New York).
• Published: 28 May 2026.

Reading scope

Full ~7-page article body read end-to-end during ingest. Six substantive contributions surfaced (24-hour-sprint model with day-shift / night-shift work decomposition; extend automation to eliminate human handoffs thesis with 30%/70% CI/CD-vs-requirements-through-coding cost decomposition; knowledge graphs as AI memory layer infrastructure prescription with the librarian agent + anti-grand-ontology design principle; team-size quantification 100 → 60 FTEs + 200 → 100 person-years + 10-role-pod → 3-role-pod; outer-loop roles integration prescription with policy-as-code; capacity-reinvestment-not-cost-cut normative framing with three-priorities operationalisation). Closing thesis on human roles concentrate at architecture + product judgment + system design + institutional knowledge + technical coherence captured. Five single-source-deferred concept-page candidates flagged: 24-hour sprint model; factory of agents; librarian agent; policy-as-code as outer-loop integration primitive; knowledge as production infrastructure. Three exhibits captured (the 24-hour timeline; the agent-squad pipeline diagram; the team-size quantification). Boundary AI cited as 2026-04-02 source for the Software development cost reference — flagged as plausible adjacent ingest target.