McKinsey's Technology Trends Outlook 2026 Adds Agentic Software Development
The report is a large document. McKinsey's PDF is dated September 2026, labeled the sixth edition, and credited to Michael Chui, Roger Roberts and Tanguy Catlin. It covers 14 technology trends that define 2026, expanding our coverage from last year to include two new fast-emerging domains: agentic software development and AI for scientific discovery and engineering. The trends fall into three broader categories: AI revolution, compute and connectivity frontiers, and cutting-edge engineering.
Agentic software development is the addition most relevant to Windows developers and enterprise IT. It means AI systems that do more than suggest code: they plan, write, test and change software with less direct human involvement. McKinsey's report page says AI agents are moving toward handling end-to-end digital tasks. It also flags the security risks that come with faster AI-generated code and notes that AI is used on both sides in cybersecurity. Four of the 14 trends are AI-specific, and McKinsey says AI underpins or amplifies the other ten.
The report's framing is that the technology story of 2026 has moved off the screen and into the physical world. It points to power grids and chips that underpin the data center boom; in the intelligent robots that embody AI; in the agentic systems discovering new chemical compounds; and in the launch pads sending thousands of satellites into orbit. The money figures are large. Energy technologies alone drew nearly $200 billion in investment in 2025, among the highest capital influx in any technology domain. And spending on AI infrastructure doubled in a single year.
Saunders gives the report credit for this. He writes that it sees what "lesser forecasters miss": AI is spreading into hardware, power distribution and skilled labor. His objection concerns what the report leaves out.
Why McKinsey's Methodology Captures Momentum but Not Control
The report's method explains a lot. McKinsey says it collected data on six tangible measures of activity: search engine queries, news articles, patents, research publications, equity investment, and talent demand. It combined patents and research publications into an innovation score, and news mentions and searches into an interest score. Scores are indexed from zero to one relative to the other trends studied. The analysis also draws on qualitative interviews with business leaders, policymakers, investors, and experts. Each trend also gets an adoption score from one (frontier innovation) to five (fully scaled).
McKinsey lists some limits of its own. It says interest scores can be inflated by deliberate efforts to generate news and searches. Its equity-investment figures leave out companies' own capital and operating spending. Its talent-demand data comes from a proprietary platform that draws mostly on English-speaking countries. The 2026 investment figures are first-half numbers projected forward to a full year. The report also says the list is meant to be useful rather than "perfect or complete."
Those measures track activity: how much money, attention and hiring a technology attracts. None of them tracks who holds decision rights over a deployed system, whether a customer can audit an automated decision, or how hard it is to move off a supplier. Governance often shows up as contracts, architecture choices and regulation rather than patent filings. A method built on the six signals above is unlikely to rank it as a "trend," even if it shapes how every other trend gets adopted. That is our inference from the method, not a claim McKinsey makes. It does suggest the gap Saunders describes comes partly from how the report measures things, whatever the authors intended.
The report does touch on constraints. One secondary summary of it, from the investment-research site Idea Farm, describes AI's expansion as increasingly constrained by power, infrastructure, cybersecurity, talent, and the challenge of converting adoption into measurable returns. The same summary cites figures that 89% of organizations regularly use AI, yet only 37% report positive EBIT impact and 93% say they exceed AI budgets. The report therefore covers friction in AI adoption. Its focus is cost, capacity and return on investment rather than who controls the systems.
Saunders' Operational Sovereignty Test: Inspect, Limit, Switch, Stop
Saunders calls the missing trend operational sovereignty. He says McKinsey treats technology control as a footnote, puts it under vague "Key Uncertainties," and frames governance as an "open question." McKinsey's summary page and the report excerpts checked for this article don't include those phrases, so they are best read as his description of the 143-page document, not confirmed quotations.
His definition is the most practical part of the column. He argues that sovereignty in an AI-driven economy means more than keeping data inside a national border. It requires authority over the systems that interpret that data and act on it. In his words, an enterprise or government must be able to "inspect automated decisions, enforce hard limits, change infrastructure suppliers and retain a clear path to human intervention — the off switch."
That breaks down into four tests an IT department can use:
- Inspection means authorized staff can see the inputs, outputs and decision records behind an automated action.
- Limits means the organization, not only the vendor, sets and enforces boundaries on what an agent may do.
- Portability means data and workflows can move to another provider or architecture without being rebuilt from scratch.
- Intervention means a human with clear accountability can pause, override or shut down a workflow.
Saunders says these fights are happening now rather than later. He describes telecom carriers trying to keep their edge networks from becoming "dumb pipes" and hyperscalers pushing to own the data center stack and custom accelerators. He argues that "the real prize is not infrastructure; it is the control layer above the infrastructure." With agentic tools now writing code and carrying out multi-step tasks inside enterprise environments, that framing matches how many administrators already experience the problem.
The "off switch" should be read as a design goal rather than a feature every system has. Deployments can span several suppliers, models and business processes, and there may be no single point where everything can be stopped. The practical lesson is to design for intervention when a deployment is planned, rather than assume it can be added later.
McKinsey's Own Numbers on Microsoft and Hyperscaler Silicon Support the Critique
The report's own data backs up part of Saunders' argument. McKinsey's summary says Amazon, Google, Meta and Microsoft are working with semiconductor firms to co-design custom silicon for AI models, and that some hyperscalers are looking at selling chips to outside customers. Vertical integration of that kind, from chip to cloud platform to model to agent, is the "control layer" consolidation Saunders describes. McKinsey reports it as an investment and innovation trend. It does not draw out what it means for customer choice.
The infrastructure constraints point the same way. McKinsey says US data centers running AI workloads alone are projected to use as much electricity by 2030 as California does today, and that transformer lead times in many markets are longer than two years. When power and equipment are that scarce, the companies that already have capacity gain leverage over the ones that need it. McKinsey also reports that five trends are on pace to receive more than twice their 2025 investment in 2026: agentic software development, AI infrastructure and model architectures, AI for scientific discovery and engineering, space technologies, and robotics. Five of 14 trends are on pace to at least double 2025 investment in 2026, led by AI-linked and space technologies.
These figures don't show that hyperscalers are seeking what Saunders calls "absolute market capture." They are McKinsey's own claims and forecasts, partly projected from first-half data. They do show that decisions about AI are becoming decisions about suppliers, compute, energy and contracts. For an enterprise running Azure, Microsoft 365 Copilot or third-party agent frameworks, those decisions determine who can change the rules later.
McKinsey's hiring data adds context on maturity. The report says more than three-quarters of job postings in its four AI trends, plus application-specific semiconductors, were in research and development. More than half of postings in connectivity, cybersecurity, energy, life sciences and mobility were in non-R&D roles. McKinsey reads this as a sign that the areas are at different stages of maturity. For governance, it suggests AI hiring is still weighted toward building the technology rather than operating and overseeing it.
Where the Fierce Network Column Goes Beyond Its Evidence
The second half of Saunders' column moves from the report to an attack on McKinsey. He calls the omission "calculated silence" that serves a client base "obsessed with absolute market capture." He then lists clients and controversies, including work for the CIA and ICE, Israel's Defense Ministry, the Saudi government, Gazprom and VEB, and congressional scrutiny of McKinsey advising the Pentagon while consulting for state-linked organizations in China.
Some of the history involves documented legal settlements. Saunders cites McKinsey's agreement to pay $650 million to resolve criminal and civil investigations into its advice to Purdue Pharma on OxyContin sales. He also cites a $122.85 million criminal penalty agreed by its South African subsidiary over a scheme to bribe officials at Eskom and Transnet, and the firm's earlier praise of Enron. That record has been widely reported. None of it explains why the 2026 report is structured the way it is. The motive Saunders assigns is his interpretation, and the report's methodology offers a simpler partial explanation: a survey built on patents, investment and hiring signals is poorly suited to tracking governance.
Two more claims in the column need narrowing. He writes that "regulators everywhere outside the U.S." are rewriting compliance, privacy and safety rules for autonomous workflows. That broad claim doesn't name specific laws or dates. And his closing advice to "consider the agenda behind the authors' narrative" applies to every trend report, vendor white paper and op-ed, his own included. He is a communications analyst who writes regularly about operational sovereignty, and Fierce Network notes that its opinion pieces do not represent the publication's views.
Other summaries of the report differ on emphasis. One, from a UK business school, says the outlook emphasises the growing importance of data governance, ethical AI deployment, and digital resilience. That summary is thin and uses generic language. Still, it shows that whether McKinsey "ignores" governance depends partly on what counts as governance: acknowledging risk, or setting out who controls the system.
What This Means for IT Teams Deploying Agentic AI
Organizations evaluating agentic software development tools, AI coding assistants or autonomous workflow platforms in the coming budget cycle should build Saunders' four control tests into procurement now. McKinsey's report is useful for deciding which technologies to watch and how much capital is moving into them. It doesn't cover vendor control, auditability or exit terms, so teams need to assess those separately. Teams still at the proof-of-concept stage have time. Teams already scaling agents into production should check these points before contracts renew.
One established framework for this work is the NIST AI Risk Management Framework. It is voluntary and not specific to any sector, and it organizes AI risk work into four functions: Govern, Map, Measure and Manage. Governance runs across the other three. The control tests below are an editorial checklist drawn from Saunders' argument and that general approach. They are not NIST requirements.
- Map each AI-driven workflow to the specific supplier, model and hosting platform it depends on, including dependencies on hyperscaler custom silicon or proprietary model APIs.
- Confirm that authorized administrators can retrieve logs of an agent's inputs, outputs and actions, and that those records stay with your organization after a contract ends.
- Name an accountable owner for every agentic workflow, with documented authority and a tested procedure to pause or roll back the agent's actions.
- Negotiate data and configuration export terms before deployment, because moving off a vendor's agent framework later is far harder than choosing one.
- Treat vendor adoption and investment figures, including McKinsey's projected 2026 doublings, as signs of market momentum rather than evidence that a tool is mature enough to operate or govern.
- Review AI-generated code with the same security scrutiny as human-written code, given McKinsey's own warning about the security risks of faster AI-generated output.
McKinsey's Technology Trends Outlook 2026 is a well-sourced map of where capital, patents and hiring are going, and it rightly shows AI moving into chips, power grids and robots. Saunders is right that the map leaves out who holds control over these systems. That gap matters more with each quarter that agentic tools spend in production. His explanation of the gap depends on accusations the report doesn't support, but the gap itself is real. As hyperscalers including Microsoft co-design their own AI silicon and AI infrastructure spending keeps doubling, the terms enterprises accept on inspection, portability and intervention in this procurement cycle will largely determine who controls these systems later.