Writing
Published here.
Most of what a review reports, the steering group half-knew already. The value is not the discovery. It is that the finding is now written down, evidenced, and impossible to un-hear.
Twelve weeks ago I argued that the bottleneck keeps migrating outward. This is the closer for that series — about what happens when world models bleed into non-robotics products and when agent systems start learning at the harness layer. The era that comes after meshes.
On August 2 the EU AI Act's main provisions became enforceable. The honest story isn't that compliance became a legal review — it's that compliance became an architectural constraint. The architecture that responds well is the agentic mesh. Here's what that actually looks like, and the order to build it in.
A 3.8B parameter model now matches GPT-4o on extraction tasks. Phi-4, Gemma 4, and Qwen 3.5 changed the calculus for what runs locally and what runs in the cloud. The shift isn't 'small models won' — it's hybrid as the default architecture. Here's the pattern that actually ships.
88% of organizations reported confirmed or suspected agent incidents last year. 45.6% still rely on shared API keys. Inside twelve months, agent identity moved from think piece to RFP line item. Here's what changed, and what good looks like now.
ARC-AGI-3 launched in March 2026. Humans score 100%. Frontier AI scores under 1%. The 99-point gap is the smaller story — the bigger one is why static reasoning benchmarks have stopped predicting whether a production agent will actually work.
Agent = Model + Harness. Mitchell Hashimoto posted that framing on February 5, 2026. Within weeks it was the field's shared vocabulary. Here's why the term stuck so fast, what a real harness actually looks like, and why this is the layer where competitive engineering work has migrated.
Three years ago the answer to 'how do I keep the value when the senior engineer leaves' was wiki pages nobody read. The answer now is the manifest file the agent reads on every run. Here's why structured skills beat fine-tuned models — and what that changes for engineering leaders.
Going from one agent to two is harder than going from zero to one. That's the orchestration problem — and the place where 2026's most interesting infrastructure work is happening. Here's the taxonomy of coordination patterns and which one to actually start with.
Claude's OSWorld benchmark score went from under 15% to 72.5% in 18 months. That isn't a research curve — it's a deployment curve. Here's what crossing that threshold actually unlocked, and which workflows are now ready for a browser agent.
97 million monthly SDK downloads. Around 10,000 servers in the public registry. 41% of orgs in production. MCP didn't win loudly — it won the way infrastructure always wins. Here's what that means for how you procure systems now.
Karpathy's definition of context engineering is doing a lot of work. The hidden shift in 2026 is that memory has detached from the rest of context engineering and become its own product category — with five distinct design camps.
GitHub Spec Kit crossed 90,000 stars in May 2026. Here's why spec-driven development isn't documentation-first — it's a contract that survives agent re-runs, and why that changes how engineering works.
Most SAFe implementations don't fail because the framework is wrong. They fail because someone drew the Agile Release Train boundaries on a whiteboard that already had the org chart on it — and never
*Most AI adoption advice optimises for the individual power user. But when one person on a squad uses Claude heavily and the rest don't, you don't get a faster team — you get a fractured one.*
Individual AI productivity rises reliably with tool access. Team-level output doesn't follow. The gap isn't a tooling problem or a governance problem — it's a ritual design problem.
*When only two or three people on a team use AI daily and the rest don't, you don't get a productivity uplift. You get a knowledge asymmetry that erodes collaboration — and no training course will fix
Most scaled agile failures are diagnosed as methodology problems. The real culprit is structural — and it's hiding in plain sight.
Most Danish enterprises still run their PMO as a reporting and compliance function — tracking RAG statuses and Gantt charts while actual business value quietly evaporates. The profession has moved on.
*Digital transformations don't fail because of strategy or technology. They fail in the unowned layer between C-suite ambition and team-level execution — and most organisations don't even have a name
Most Nordic enterprises have adopted agile delivery methods but left their annual budgeting and governance cycles completely intact. The result is governance theatre — and it's the single biggest reason agile transformations fail to deliver at scale.
Nordic enterprises are delegating EU AI Act compliance to legal departments and IT security teams. They're solving the wrong problem — and the August 2026 deadlines won't wait.
The returns gap is not a strategy problem or a technology problem. It is a programme governance problem — and it is solvable.
Most Danish enterprises are treating EU AI Act compliance as a legal or IT project. That's a structural mistake — and the August 2026 deadlines will expose it. Here's what board-level AI accountabilit
Ambient computing is dissolving the boundaries between digital and physical worlds. Discover how invisible technology is quietly transforming how we live, work, and interact.
Most Danish mid-to-large enterprises have run successful AI pilots. Fewer than 20% have operationalised AI at scale. The gap isn't technical — it's organisational. Here's how to close it before the EU AI Act enforcement window hits in August.
The average PMO gets restructured every two years. Yet recent research shows PMO roles explain nearly 73% of variance in strategic plan execution. Here's how to bridge that gap.
Why law firms that experimented early with AI—despite failures—built invisible assets positioning them for tomorrow's competitive edge.
How AI is reshaping legal services by automating routine tasks, enhancing research, and enabling data-driven insights.
Quantitative frameworks for measuring ROI and value realization in AI programs with clear metrics and benchmarks.
Government AI adoption success factors: lessons learned and practical frameworks for public sector implementation.
Key success factors for digital transformation: culture, collaboration, and continuous learning approaches that drive results.
Essential principles and practices for building high-performing agile teams in digital transformation contexts.
How strategic foresight helps organizations anticipate disruptions, craft resilient scenarios, and shape digital strategies.
Human-centered change management strategies for successful AI adoption, addressing cultural transformation challenges.
How AI-enabled knowledge management systems capture, organize, and leverage collective intelligence in professional services.
Building comprehensive AI governance frameworks using NIST, ISO 42001 standards, and EU compliance requirements.
A Scrum Master's guide to building high-performance agile teams through psychological safety and collaboration.
How to lead AI initiatives in law firms through structured programs, governance frameworks, and cross-functional collaboration.
Strategic approach to digital transformation in public sector agencies, focusing on service delivery and operational efficiency.
Lessons learned from SAFe implementation: how to successfully scale agile methodologies in enterprise organizations.
From theory to practice: implementing AI and machine learning solutions in municipal services with real-world examples.
How modern PMOs drive strategic value by adapting to digital transformation and agile delivery methodologies.
How platform modernization enhances customer experience through improved digital touchpoints and service delivery.
Thirty minutes. I will confirm whether I can act.
Arrange a consultation