Skip to contentLatestThe End of Software Built by StrangersPeople who do not understand the work are still choosing the software everyone else must use. The proxy tax is the cost of that distance.Oct 2026All writing
- Oct 2026The Entrance Exam Is Gone.The tools did not just make existing players faster. They let people who were never in the industry enter the market. Engineering is no longer always the prerequisite for having a product. The market is replacing the institution as the entrance exam.
- Oct 2026Anonymised Feels Comfortable. It Is Not.Resequencing IDs so a European dataset looks abstract is not anonymisation. GDPR still calls it personal data. The comfort ritual retires the people who used to worry and hands the model a biography. Put the permission on the tool.
- Oct 2026The Ticket Queue Was a Rationing SystemThe analyst liturgy (ticket, steward, warehouse, dashboard, slide) rationed a world in which querying was expensive. AI-native work is a conversation with the estate in seconds. The constraint is giving them that freedom without lying about people.
- Oct 2026Share the Taxonomy or Do Not BotherAn agent cannot be AI-native on fourteen definitions of the same person. Shared taxonomy is the contract: entity, grain, owner, retention, jurisdiction, loaded like a skill. Skip it and every tool becomes a private dialect.
- Oct 2026Graph and Vector Are Tools, Not a ReligionVector finds what feels similar. Graph knows what is the same person, and which join is allowed. Expose both through MCP under one taxonomy and one kill switch. The storage bake-off is the wrong argument.
- Oct 2026The Bureaucracy Is About to Become AutonomousWe are teaching machines to execute decisions nobody remembers making. AI may not eliminate bureaucracy. It may industrialise it. Decision debt, provenance, and controlled unlearning are the response.
- Oct 2026Charter the Lab: When to Fund It, When to Kill It, What Legal SignsA lab is worthwhile only as a foundry. Fund it when you can name four operators, an hour-scale witness, one allowed data slice, and a board that can sunset in public. Otherwise skip the glass and run FDE in the business.
- Oct 2026Establish the Lab: Hunt Change-Makers and Give Them the Tools NowThe idea-to-execution gap has collapsed. A lab is not innovation theatre, it is the empowerment system behind FDE: hunt the operators already pulling, put a harness in their hands this month, and convert irreplaceable talent into loops that graduate.
- Oct 2026The FDE Frontier: Designing Forward Deployed Teams That CompoundEnterprises copied the Forward Deployed Engineer title and missed the operating system. FDE is not SWAT, staff-aug, or CoE theatre, it is a triad at the edge: operator, builder, platform. Design residencies with exit conditions, ship harness not decks, and measure the loop.
- Oct 2026Skills Left the Vendor: When the Folder Became InfrastructureOn 22 Oct Anthropic donated Agent Skills to the AAIF, 70+ runtimes, a million published skills. A letter can still kill a model. It cannot as easily kill a folder the industry agreed to load. Canonical tree, dual-home runtimes, hunt authors not vendors.
- Oct 2026The Digital Thread Is a Harness, Not a SloganVendors sell a digital thread. Most are slideware between PLM, MES, and a twin nobody trusts. Build it as a skill graph with owners, evals, identity, and a kill switch, cathedral on the thread, tangents allowed to die. Software FDE and hardware FDE are one OS.
- Oct 2026Your Hardware Talent Was Always the BottleneckImplementation in CAD, firmware, and test is compressing. Hardware orgs still ration a priesthood. The scarce asset is the IC on the bench: domain, taste, qualification memory. Hunt them, give them skills and a floor, filter for the mindset that authors a skill from a failed loop.
- Sep 2026Air Gap, Open Weights, and the Plant That Cannot Call the APIOT, ITAR, and air-gapped sites cannot depend on a frontier API. Boards still buy Cloud Copilot for those cells. Run portable skills and on-prem inference behind the fence, Market B with a safety case. Workflow ownership on disk is the hedge.
- Sep 2026Cathedral Machines, Tangent FixturesAgents spawn CAD, HMIs, and twins at prompt speed. Hardware culture treats every surface as capital. Two speeds: certified product and regulated HMI are cathedrals; fixtures, jigs, and Tuesday twin views are tangents. Label at birth. Assign taste. Sunset in public.
- Sep 2026The Lab Next to the LineIndustrial innovation labs are usually a tour. A lab that empowers sits next to the cell: hunt unofficial scripters, put a role-bundle of skills in their hands this month, graduate / library / sunset every loop, keep a stop a technician will hit.
- Aug 2026Forward Deployed on the LineIndustrial AI is hiring FDEs into plants, CAD, and robotics. Enterprises will copy the title into OT as SWAT. Design residencies against yield and cycle time: process owner + FDE + platform. Ship skills and evals a technician can keep. Gravity version of the FDE OS.
- Aug 2026Evals When the World Can BreakA 90% pass rate is a model card in software and a recall if the 10% is a weld. Score a ladder from model to sim to bench to line. Fail loud. Cite materials. Put the kill switch in the path. A skill without an eval is a style guide.
- Jul 2026Skills Became InfrastructureAgent Skills are an open folder standard every serious runtime loads. Progressive disclosure, evals-first authoring, organic marketplace, role bundles, portable paths, and a physical first library as proof the same object survives the bench. Stop collecting prompts in Notion.
- Jul 2026The Last Mile Was Always AtomsSoftware learned a model can vanish on a Friday. Hardware never believed otherwise. The last mile was always geometry, pinouts, toolpaths, and the machine that still faxes. Apply the same OS, skills, evals, FDE, lab, not a Physical AI CoE. Gravity showed up.
- Jun 2026After the Kill Switch: What the Trench KeptTwo weeks after Fable went dark, the question was not fairness. It was what you still had on disk. Evals, routers, open-weight forks, skill folders, not the apology. Rent the horse. Own the harness. Capability you cannot copy off the laptop was never yours.
- Jun 2026Fast-Fashion Interfaces: Assigning Taste When Every Surface Is TemporaryMost enterprise UI is becoming tangential, spawned for a user and a request, then replaced. Cathedrals still matter; arbitrators of taste must govern what compounds. Code-editor logic meets org design: clone, invert, remix, sunset.
- Jun 2026Nerfed Frontier, Open Weights: What the Trenches Actually Build OnFrontier APIs are policy-gated and discontinuous, Europe will keep buying nerfed tiers while boards sleep. Serious builders fork open weights, tune on their data, and share harnesses in the trench. Portability wins when policy gets loud.
- Jun 2026The Fable Directive: When Model Wars Go GeopoliticalAnthropic recalled Fable 5 and Mythos 5 worldwide after a US export-control directive, triggered by a narrow jailbreak class the lab said is widely available elsewhere. Frontier deployment is now geopolitics: kill switches, nerfed access for allied markets, and labs becoming integral to US economic infrastructure.
- Jun 2026The Harness Is Becoming More Important Than the HorseFoundation models are strong enough for serious operational use, but most teams still measure progress by model choice, not by whether intelligence survives contact with production. The durable advantage is moving to the harness: interface, memory, permissions, workflows, guardrails, and the ops layer that makes models useful every day.
- Jun 2026The Corporate Impotence CycleThe people with the most authority often have the least understanding of what is actually happening, not because they do not care, but because leadership consumes summaries of the business instead of experiencing it. Detachment breeds poor decisions, more reporting, and greater distance from the work. The fix is proximity, not another transformation programme.
- Jun 2026I Built a Reddit GTM Engine Because Launching Products Was Taking Too LongEvery launch required the same repetitive work, research, references, assets, copy, variations. Creative Patterns discovers winning patterns; We Go-To-Market turns them into launch-ready assets. Human stays creative director; the system handles production. Most AI products generate outputs, I became more interested in generating entire go-to-market systems.
- May 2026Paperclip and the Emergence of Organisational AIMost AI tools are designed around a single user. Companies are networks of responsibilities, approvals, handoffs, and budgets. Paperclip mirrors organisational structure, org charts, goals, heartbeats, governance, and orchestrates agents inside that shape. An early glimpse of digital organisational design, not another agent framework review.
- May 2026Dealing With AGI: Tracing the Singularity Back to TodayA reverse-traversal from the singularity back to today, layer by Maslow layer. Which professions dissolve, which compound, what the model builders owe, and what serious people should be doing with the next two years. The written companion to the dealingwithagi.com immersive piece.
- May 2026In-House Excellence vs Agency Fuel: Why Brand Acceleration Needs BothIn-house build capacity is real inside brand teams. The right read isn't "agencies are dead." It's that in-house excellence is the maturity signal that separates the agencies worth keeping from the ones built on headcount arbitrage. The new contract: brand teams own the production loop; agencies supply transformation, taste at scale, reality, and concentrated problem-solving.
- May 2026Fast-Fashion Software: Disposable Tools, Compounding TasteMost software an operator ships in the next decade should be disposable. Fast-fashion software is the natural operating mode of the IC age. The compounding part is the layer underneath, skills, knowledge, taste, brand, not the artifact.
- May 2026The Skill Library: 20 Composable Capabilities Every AI-Native Team Should Ship FirstIf skills are the reusable unit and agents are the orchestrator, the next question is which skills to write first. Twenty named, testable capabilities across writing, visual, research, video, and engineering, the IC-grade operating layer that compounds.
- May 2026Hot Work and Deep Work: Two Modes of an AI-Native OperatorMost teams pretend there is one mode of work. There are two. Hot work brings the power; deep work earns it. The teams that name the two modes, build a cadence between them, and refuse to silently blend them produce the durable, defensible work of the next decade.
- May 2026The Battle of Mindsets: When Everyone Has Claude Code, What Filters Talent?When the tool is in every hand, the filter moves up the stack to mindset, framing speed, cheap falsification, taste under abundance, and ownership of the loop. The 2026 hiring loop will be unrecognizable in three years, and the right gamified evaluation/training surface is wide open.
- May 2026As Fast As You Can Think: Closing the Conception-Execution GapFor most of computing history, the bottleneck was translating thought into syntax. The keyboard always lost something on the way out. What replaces it is a system that infers the missing half of your sentences, and the operators who already knew what they were trying to do are about to look superhuman.
- May 2026The Individual Contributor Age: When Engineering Capacity Stops Being the BottleneckFor two decades, software's most expensive resource was the engineer. AI compresses engineering capacity to near-zero, and the bottleneck moves to the person who knows what is worth building. The IC age is structural, not tactical, and it rewires the org chart underneath it.
- May 2026More Variants Are Cheap. Creative Learning Is the Bottleneck.Ad platforms generate more than ever, but most teams drown in exports. The bet behind Creative Patterns: the future of paid is owned creative intelligence, portable memory, not more noise, aligned with the machine-readable design thesis from DESIGN.md and Stitch.
- Apr 2026The Portable Loop: Import, Enrich, Patterns, Runbook, ExportHow Creative Patterns works: not another chatbox that writes ads, but a layer above the generators, CSV plus context, scored patterns, LLM-ready runbooks, and exports you can bring back to any platform. Same compounding bet as machine-readable design.
- Apr 2026When Production Is Free, Taste, Memory, and the Loop CompoundsWhy the winners are not the teams with the most AI variants, they are the teams with the tightest creative feedback loop. Extraction, auditability, and the same defensible work as high-stakes design systems, applied to paid learnings you actually own.
- Apr 2026A Public /writing: Ownership, Reputation, and a Machine-Readable Home for the ThesisWhy a canonical writing surface on the open web beats scattering notes across feeds. Ownership of narrative, compounding trust, and a spec humans and systems can read, tied to agentic infrastructure, not to someone else’s algorithm.
- Apr 2026Ship the Note: When Essays Are Product, not PerformanceWhat publishing here actually does for the build: case studies, hot takes, and operating discipline, Strategy, Architect, Build in prose, with receipts that back the rest of the site.
- Apr 2026DESIGN.md, Stitch, and the Bet on Machine-Readable BrandDESIGN.md, Brand Lockup, Stitch, Creative Patterns, and a blunt bet: when pixels are cheap, own the spec, audit, and the infrastructure that can say no.
- Apr 2026Build Skills, Not Agent Armies: Thin Orchestration and Why It CompoundsAgent swarms look like progress until you inherit a second org chart of prompts, retries, and debug sessions. A SCQA-grounded take: invest in reusable, testable skills first, then let a thinner agent layer coordinate. How Anthropic's Agent Skills framing maps to safer, cheaper operations.
- Apr 2026Personal LLM Knowledge Bases: Strategy, Capture, and Why PageStash Belongs in the StackKarpathy's LLM wiki pattern turns reading into a compounding asset, not dead bookmarks. A five-move framework for ingest, compile, operate, reinvest, and hygiene, plus where PageStash substitutes for Obsidian Web Clipper on the web-to-structured-text path.
- Mar 2026n8n and the Cautious Operations Playbook: Innovation Labs, Enterprises, and Social ImpactWorkflow automation that wraps LLMs in guardrails and legacy integration, not another bespoke model, is how most orgs get to agentic ops. Practical patterns for innovation labs, enterprises, and social enterprises.
- Mar 2026LangGraph on the Ground: Durable State, Humans, and Observability for Real OperationsWhen agents touch money, liability, or compliance, demo-grade scripts fail. LangGraph's checkpointers, interrupts, and LangSmith traces are how teams ship operations that survive restarts, audits, and humans in the loop.
- Mar 2026The End of the Monolithic Product OrgWe broke monolithic code into services and sharpened boundaries with DDD. The bottleneck moved to the center. Decentralized delivery with real governance beats one product org gating every workflow.
- Mar 2026The Karpathy Autoresearch Pattern: AI That Runs Experiments While You SleepAndrej Karpathy's autoresearch hit 30K stars in a week. The program.md pattern, humans set goals, agents run experiments overnight, generalizes beyond ML to any domain with measurable objectives.
- Mar 2026Enterprise Auto-Research: Multi-Agent Intelligence That Compounds OvernightEDR, DOVA, and the shift from single-agent loops to multi-agent orchestration. Enterprise deep research is here. The organizations that deploy it first own the information advantage.
- Mar 2026NVIDIA Full Throttle: The Enterprise AI Stack Everyone Must OwnAt GTC 2026, Jensen Huang announced a thesis: every enterprise must own its AI stack, and the best economics come when that stack is NVIDIA, end to end. GPT-OSS on NIM, NemoClaw, and the five-layer cake.
- Mar 2026Stop Wasting My Tokens: The Anti-Verbosity ManifestoTokens are the currency of the Agentic era. Learn why "polite" LLM filler is a tax on your latency, your margin, and your intelligence-density.
- Mar 2026Beyond JSON: High-Density Data for Agentic PayloadsFormatting overhead is structural waste. Explore how TOON and structured Markdown maximize entropy in your agent-to-agent swarms.
- Mar 2026The Efficiency of Hostility: Why I Swear at My ModelsLLMs aren’t human. They don't have feelings. But they do have semantic pathways triggered by high-stakes emotional language. Here is why being a "dick" actually gets you better results.
- Mar 2026Beyond Agile: The Rise of the Micro-TeamAgile was built to manage the friction between idea and delivery. When AI collapses that friction to zero, the methodology becomes the bottleneck.
- Mar 2026The SMB Arbitrage: How Small Teams Out-Execute through AutomationWhy the real winners of the agentic revolution aren't enterprises, but agile small businesses rewiring their operations.
- Mar 2026The Fractional Digital Employee: Scaling Operations with AgentsStop treating automation as an IT project. Treat it as a hire. How small businesses are reclaiming hours.
- Feb 2026Navigating the Agentic Stack: A 2026 Framework Selection MatrixEngineering advice for companies choosing between LangGraph, CrewAI, Pydantic AI, and the managed walled gardens.
- Feb 2026Defensive Agents: The Weaponization of AdoptionThe rise of the defensive layer: how model providers are leveraging AI-powered anxiety to drive enterprise adoption.
- Feb 2026Hedging with Open Source: Models, Data, and AgentryAvoiding the walled gardens: why local models and enriched open-source data are the secret to scalable agentic systems.
- Feb 2026Agentic Marketing: The End of the FunnelWhen the buyer is an agent, the marketing funnel collapses into a data-driven satisfiability problem.
- Feb 2026The Role of Personal AgentsFrom static UIs to intent-driven swarms: the shifting landscape of how we delegate our lives to machines.
- Feb 2026Why I Back Agentic InfrastructureThe pattern across Siren, Skydio, Swan, and Webflow: systems that scale intelligence and autonomy.
- Feb 2026The Agentic EnterpriseWhat Dreamforce 2025 made clear: the divide is accelerating, and brand authenticity is the defensible moat.
- Feb 2026Design Systems as CodeAdvertising creatives shouldn't fight tools. They should have systems that behave like code.
- Feb 2026Brand Lockup: Design Systems for AdvertisingProblem, approach, outcome, how we built design systems that behave like code.
- Feb 2026ThreatBase: Agentic Intelligence for OSINTScaling open-source intelligence with agentic workflows, what we built and why.
- Feb 2026PageStash: Capture, Structure, RetrieveTurning the web into a queryable knowledge layer for research and intelligence.
- Oct 2025Dreamforce 2025 takeaways↗The "agentic enterprise" is no longer theoretical. Five critical trends on AI, Salesforce's stack, labor transformation, and brand authenticity.
- Jul 2023The job of being a creative just got a lot more specialized↗How RAPP leverages AI, tech, and data into campaigns. Interview with Michael Couch on precision marketing and evolving creative roles.
- Digiday Programmatic Marketing Summit, Google's Cookieless Solutions↗Speaking with Digiday on Google's cookieless solutions and adoption among advertisers.
- Sugaring the Pill: How to Modernize Pharma Marketing↗Content automation, MLR workflows, and modernizing pharma marketing operations. Co-authored with David Whitmore and Jennifer Corr.