The older you get, the more you realise that most of the time, people are making things up as they go along. Stumbling forward with rules of thumb and mental models and “experience.” You pretend you knew all along either way.
AI’s impact on the workforce is the current case. I count nine steelmanned theories so far. Each one felt right until the next one landed.
Nine steelmanned theories about AI and work, each upending the last
The sequence of predictions — roughly chronological, 2023–2026
Source: barnabyrobson.org. Nine steelmanned theories from AI workforce discourse, 2023–2026.
Anthropic’s own data puts numbers to the confusion. Their Economic Index maps theoretical AI capability against observed usage by occupation. Theoretical coverage reaches 0.8 across management, business, finance, computing. Observed usage barely touches 0.3 in any category. Everyone is still figuring it out.
AI could theoretically perform most knowledge work — actual usage is a fraction of that
Share of job tasks by occupational category, 0–1 scale
Source: Anthropic Economic Index (2025). Theoretical capability from Eloundou et al.; observed from Claude conversation data. Values approximate — read from original figure.
Some of the corporate communications on the coming change have been shockingly bad. Matthew Prince, CEO of Cloudflare, laid off more than 20% of his workforce and labelled the people he cut “measurers.” Chamath Palihapitiya on the All-In Podcast: “You could not have written a worse memo. You reduce humans to a label called the measurer and then you lay off all the measurers. You put a scarlet letter on their back.”
Meanwhile Zuckerberg laid off 8,000 people while simultaneously installing recording software on the survivors’ computers to study their work and train AI models. Some of those employees had built AI tools during company hackathons to make their jobs more efficient. Then they got cut.
Aaron Levy’s views around “AI washing” resonate: CEOs are “uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen.” They see the demo. They never see the deployment.
My mind has changed too.
I used to think SaaS was finished. I’ve moved. Systems of record — the Salesforces, the ServiceNows — probably survive. The value-added services layer sitting above those platforms — that’s where the trouble is.
What’s actually happening is that AI creates more work. Dan Shipper’s essay “After Automation” makes this case well — AI commoditises the residue of human expertise, collapses the value of default output, and creates demand for what’s different. David Sachs (former US crypto and AI czar) on X: “We’re already seeing a 14x YoY increase in GitHub commits, and it’s accelerating. AI has dramatically lowered the cost of writing code, so it’s now being used across far more businesses, applications, and use cases. We’re at the beginning of a massive productivity boom.” The fact that coding — AI’s breakout use case — has increased demand for software engineers should call into question the entire mass job-loss narrative. Software engineer job postings are surging while overall postings stay flat.
Software engineer job postings surged while overall postings stayed flat
Indeed job postings, 21-day moving average. Indexed to Jan 2024 = 100.
Source: Citadel Securities, Indeed. Software engineer postings (Indeed index, LHS in original) and overall Indeed job postings (RHS in original), normalised to Jan 2024 = 100. Figures are for illustrative purposes only.
GitHub’s own data confirms the acceleration. Commits reached 1.4 billion per month — up 7x in three years. Pull requests and new repositories show the same inflection. All three went vertical from December 2025 when agentic development workflows arrived.
GitHub commits hit 1.4 billion per month — up 7x in three years
Monthly commits, 2023–April 2026. Agentic workflows accelerated sharply from December 2025.
Source: GitHub blog, “An update on GitHub availability” (28 April 2026). Pull requests merged (90M/month) and new repositories (20M/month) show the same inflection.
AI is not taking your job. Someone using AI more effectively than you will. It’s on your employer to train you — context architecture, evaluation frameworks, blast-radius management. And it’s on all of us to spend our own time honing our skills. This is the non-negotiable.
Cope
But orchestrating AI gets complex fast. As I explored in my last article, it’s easy to start building loads of crazy shit with AI — skills, agents, cron jobs, hooks — and then drown in it. You need a mental model to link your thinking to your workflows before the whole thing spirals into a forty-skill pile of competing instructions.
The sandwich
A memorable model is what Every.to calls the “human-AI sandwich.”2Kieran Klaassen / Every.to — “You’re the bread in the AI sandwich.” From Dan Shipper’s “After Automation” (Every.to, 2026). The model: human sets the frame and parameters, AI collapses the task, human reviews, corrects, and decides what ships. You frame the task. The AI collapses it. You judge and extend the result. The human is the bread on both ends.
The human-AI sandwich
You’re the bread on both ends
Concept: Kieran Klaassen / Every.to — “You’re the bread in the AI sandwich.” Adapted from Dan Shipper, “After Automation” (2026).
I think COPE is the best mental model for structuring this: Capture, Organise, Process, Execute.1COPE draws on David Allen’s GTD and Tiago Forte’s PARA. I’ve reframed both for the agentic world — the AI handles the bulk of each phase, the human sets the frame and judges the output. Also: cope. Because that’s what we’re all doing.
The COPE framework
Capture, Organise, Process, Execute. Also: cope.
© Barnaby Robson 2024. COPE draws on David Allen’s GTD and Tiago Forte’s PARA, reframed for the agentic world.
I’ve taken this further and built what I call bstack — the b stands for Barnaby, because apparently that’s the kind of person I’ve become — a personal skill stack that encodes my workflows into skill families.3The name rips off Garry Tan’s gstack. The seven families: bknow (32 skills — knowledge management), bvoice (14 — publishing pipeline), bdeals (12 — M&A deals), bvest (12 — investing), brun (7 — runtime health), blook (7 — charts and visual output), plus 27 cross-cutting tools. Each skill encodes workflows, child skills, JavaScript and Python queries — process rather than content. 111 skills across seven families, each encoding domain judgement into reusable markdown documents — code, prompts, and rules that compound over time. They’re portable across models and tools. I may write about the bstack family in due course. It’s pretty cool.
Many organisations (ahem!..) are stuck with multiple tools from different vendors — different logic, different workflows, none of it hot-swappable across models.4The pattern I see most often: fourteen separate AI chatbots, each built for one task, no shared context layer, no way for one tool to learn from another. The organisations that build end-to-end will compound. The demo-builders won’t. All of it becomes outdated the moment the next model ships. Impressive demos. Zero compound returns.
What people should be building: specific skills and code that outlast model refreshes, iterated in real time the way engineers push code. Garry Tan frames these as method calls — markdown as the programming language, human judgement as the runtime. Skills are portable across models and vendors. When the model changes, the skill still works.
Building your own
At a personal level, skills embed your logic, your personality, your ways of thinking. They transfer across personal and professional domains. Everyone is still making it up. The theories will keep flipping. The one thing you can control is whether you have a system for coping with the noise.
Notes
- COPE draws on David Allen’s GTD and Tiago Forte’s PARA. I’ve reframed both for the agentic world — the AI handles the bulk of each phase, the human sets the frame and judges the output. ↩︎
- Kieran Klaassen / Every.to — “You’re the bread in the AI sandwich.” From Dan Shipper’s “After Automation” (Every.to, 2026). The model: human sets the frame and parameters, AI collapses the task, human reviews, corrects, and decides what ships. ↩︎
- The name rips off Garry Tan’s gstack. The seven families: bknow (32 skills — knowledge management), bvoice (14 — publishing pipeline), bdeals (12 — M&A deals), bvest (12 — investing), brun (7 — runtime health), blook (7 — charts and visual output), plus 27 cross-cutting tools. Each skill encodes workflows, child skills, JavaScript and Python queries — process rather than content. ↩︎
- The pattern I see most often: fourteen separate AI chatbots, each built for one task, no shared context layer, no way for one tool to learn from another. The organisations that build end-to-end will compound. The demo-builders won’t. ↩︎
