This Week in Work

Orchard-Claw is a small idea with a large shadow. Microsoft Research’s new paper, “Orchard: An Open Framework for Scalable Agentic AI,” names it almost in passing — one of three example agents built to prove out an open training rig, alongside a code-repair agent and a web-browsing agent. Orchard-Claw does something plainer than either of those: it manages email, calendars, and the search for information across the tools a person actually uses at work. After training on just 200 synthetic tasks, its success rate under one evaluation harness moved from 18.6% to 51.5%.

The framework around it is, on its own terms, a generous piece of engineering. Building a capable AI agent has typically required infrastructure most researchers never see — custom sandboxes, closed training pipelines, proprietary datasets locked inside a handful of labs. Orchard opens that up. It lets a team spin up thousands of isolated environments on ordinary Kubernetes infrastructure, and — this is the clever part — train an agent inside the same deployment harness it will actually run in, rather than a simplified stand-in that behaves differently once the training wheels come off. The authors call trained agent behaviour a “persistent asset”: something future systems can inherit and extend, rather than relearn from nothing each time. It is, genuinely, a useful contribution to how agents get built.

But sit with what Orchard-Claw was built to do. Not “write code.” Not “browse a website.” Manage an inbox. Protect a calendar. Chase down the meeting brief nobody sent. That is, in most organisations, a named job, done by a named person, and it is rarely the person in the room when the roadmap gets decided.

There is a version of this that looks like relief. Give the agent the first pass — the routine reschedule, the “can you find that email from March” — and let the person who currently does that job spend the freed hours on the parts that actually need a human: reading the room before a difficult meeting, knowing which “urgent” email is not, managing the version of the calendar that exists in someone’s head and not in Outlook. Executive assistants have always done more judgement work than their job title admits. An agent that clears the mechanical layer could, in principle, make that judgement work more visible, not less.

There is another version that looks like nothing at all, until the contract isn’t renewed.

The distribution of technological benefit follows existing power gradients, not technological inevitability, and Orchard doesn’t decide which version an organisation gets. Nothing in the paper does. That decision gets made afterward, in a budget meeting, by someone weighing a headcount line against a credits line — and the paper’s own numbers, 18.6% rising to 51.5% in a matter of months, are precisely the kind of number that makes that meeting easier to have.

The distribution of technological benefit follows existing power gradients, not technological inevitability.

None of which is an argument against building the thing. Open infrastructure that any lab can inspect and reproduce is better than the same capability locked behind three companies’ APIs — that much is a fair trade. The argument is narrower: the paper measures a success rate. It does not measure who currently holds that job, what they know that a benchmark doesn’t capture, or what their organisation intends to do with the extra 33 points. That measurement doesn’t exist yet, because nobody has been asked to make it, and it should exist before the decision does. The technology is unwritten. The org chart it lands on is not — yet.

Microsoft Research released its New Future of Work Report examining how AI is reshaping labour across sectors. The picture is not the utopian vision of human flourishing we've been promised. Instead, it reveals something more familiar: the distribution of technological benefit follows existing power gradients.

The research is clear on the mechanism. High-autonomy knowledge workers — researchers, designers, strategists — are using AI to expand what they can do. A product manager uses it to explore design options faster. A strategist uses it to test scenarios. An engineer uses it as a thought partner. These workers maintain control: they direct the tool, evaluate its output, keep the judgement call. AI multiplies their capability.

Frontline workers experience a different technology entirely. A warehouse associate whose picks are monitored by AI-powered computer vision doesn't feel augmented. A customer service agent whose calls are analysed for script compliance doesn't feel empowered. A care worker whose rounds are timed to the minute doesn't feel supported. The same technology that expands autonomy for some becomes, for others, a system of continuous verification.

The distribution of technological benefit follows existing power gradients, not technological inevitability.

This matters because it's not inevitable. It's a design choice. Microsoft's researchers are careful not to blame the technology itself. AI can be designed to support frontline workers — to flag unusual situations, to reduce routine cognitive load, to surface patterns humans might miss. Some organisations are doing exactly that. But many more are using it to monitor, to optimise, to tighten. Why? Because it's cheaper than trusting workers. Because it transfers risk away from management. Because it can.

The question for any organisation is whether AI expands or constrains the autonomy of workers. Not all workers. Your workers. The ones you depend on. A delivery driver using AI routing that adapts to road conditions, weather, and package type has been given better information. A delivery driver whose route is immovable and whose pace is monitored against an algorithm-derived benchmark has been handed a boss that never sleeps.

The research identifies the choice point: during implementation, when an organisation decides what the system will do. Will it augment or surveil? Will it distribute understanding or concentrate control? These decisions aren't made by algorithms. They're made by people in meetings, deciding what kind of workplace they want to build.

Which raises the question that should matter most: if we're building workplaces where AI supports human autonomy for some and constrains it for others, are we building the future of work we actually want — or just automating the inequalities we already have?

There's a particular kind of passivity that settles in when any powerful technology becomes widespread. It feels as if the technology is doing something to us — that we're caught in its logic, swept along by historical forces, passengers in a vehicle that's already moving. This is the feeling that surrounds AI now. It's coming. We adapt or get left behind.

But this is where the Microsoft researchers — Chief Scientist Jaime Teevan and researchers Jenna Butler, Jake Hofman, and Rebecca Janssen — want to intervene. They've spent months documenting how AI is reshaping work, and their conclusion is almost radical in its simplicity: we are not passengers. We are steering. The question is whether we're steering deliberately or by default.

The distinction matters enormously. Default steering — the kind most organisations do — means adopting AI tools off the shelf, integrating them into existing workflows, letting the vendor's roadmap become your roadmap. You follow the capability. You adapt your work to fit the tool. You tell yourself this is just how the technology works.

Deliberate steering is different. It starts with a question about your actual work. What do your people spend too much time on? What requires judgment they're not getting to because they're buried in routine? Where does exhaustion come from? Only then do you ask: how might AI help here? What would genuinely supportive design look like?

Most organisations follow the technology. Deliberate ones ask what their workers actually need.

The podcast, which accompanies the written report, gives concrete examples of what this distinction produces. One case: a research institution that wanted to use AI to accelerate literature review. Default approach would be: here's an AI summarisation tool, use it on your databases. But deliberate design asked: what's hard about literature review? And the researchers realised it wasn't reading summaries. It was having to trace ideas across papers, seeing connections that are only visible after you've absorbed dozens of sources. So they designed an AI system that maps conceptual relationships and helps researchers navigate the intellectual landscape. Same technology, radically different intent.

This is true across domains. A call centre could use AI to monitor agents for script compliance — the default approach. Or it could use AI to flag unusual customer situations that need experienced judgment — deliberate design. A warehouse could deploy computer vision to measure productivity per minute. Or to highlight patterns in returns that suggest design problems. The tool is the same. The choice is about whose interests the deployment serves.

What's striking about the researchers' position is that they don't present this as moral purity versus corporate evil. This is practical. Organisations that deliberately design AI deployment around what their workers actually need end up with better outcomes — less turnover, higher engagement, faster innovation. The workers aren't fighting the tool. They're using it to do their work better. That's not virtue. That's strategy.

But it requires something that default adoption actively prevents: time. Space to ask the right questions. Willingness to listen to frontline workers about what's actually hard. Permission to design something bespoke rather than buying something off-the-shelf. Most organisations don't have that permission. The pressure is toward speed, toward matching competitors, toward "digital transformation" as an abstract noun that sounds like progress.

Which raises the real question: what would it take for your organisation to shift from following technology to steering it? Not because the research says you should. But because the work you actually do requires it.

The question sounds abstract, academic. Will machines ever be intelligent? But ask it in a workplace at the moment the leadership team is deciding whether to automate a role, and it becomes entirely practical. Because the answer changes everything.

Subutai Ahmad and Nicolò Fusi, researchers at Microsoft, have been comparing what transformer-based AI systems do with what the human brain does. They're not trying to match hype with caution. They're asking a genuinely technical question: what are the differences in how these systems actually work?

The distinction they're drawing is between specific capability and general reasoning. Modern AI is genuinely extraordinary at specific tasks. Show it thousands of examples of something, and it will recognise patterns you might miss. Given a context — some words, some images — it will predict what comes next with unsettling accuracy. If the task is pattern recognition within a defined domain, AI can match or exceed human ability. This is not hype. It's real.

But general reasoning is different. Genuine intelligence — the kind humans do — involves understanding. It means grasping the underlying principles that make something work. It means adapting to genuinely novel situations you've never encountered before. It means asking yourself whether your reasoning is sound. It means recognising when the rules have changed and adjusting.

Most of what AI does is extraordinary pattern recognition. Much of what humans think is important is reasoning in the face of uncertainty.

The researchers are careful here. They're not saying machines can't get better at reasoning. They're saying that what they do now — even the most sophisticated systems — doesn't look like reasoning. It looks like the recognition of increasingly subtle patterns. And those are different things.

This matters for work because not all work is pattern recognition. Some is. Diagnosing a specific disease from symptoms and test results is pattern matching against medical literature, and modern AI is genuinely useful here. Reviewing contracts for standard risk clauses is pattern matching, and AI can accelerate this. Predicting customer behaviour from transaction history is pattern matching, and AI does it well.

But designing something new requires reasoning. It requires understanding why things work the way they do, not just predicting what usually happens. Negotiating a complex deal requires navigating genuine uncertainty where past patterns might not hold. Leading a team through change requires understanding people, not predicting their responses. Teaching requires reasoning about how someone actually learns, not just delivering content. Diagnosing an edge case — the patient whose symptoms don't fit the pattern — requires reasoning.

This is the distinction the researchers want you to sit with. As organisations implement AI, the question isn't: what's the technology capable of? It's: what kind of work is this, really? Is it pattern recognition or reasoning? Are we automating something that looks most like what these systems actually do, or are we delegating something that requires capacities these systems don't actually have?

The risk is obvious. You can deploy pattern-matching systems into reasoning-dependent work, and for a while it feels fine. The patterns hold. But then you reach the edge case, the genuinely novel situation, the moment where the underlying logic changed and the patterns broke. And you discover that you've automated away the capacity you actually needed.

The researchers aren't arguing against AI. They're arguing for precision about what it is and what it isn't. Treat it as a pattern matcher and it's extraordinary. Treat it as reasoning and you're misusing the tool and potentially putting human judgment where you shouldn't.

Which means the urgent question for any organisation is not: should we use AI? It's: do we understand exactly what kind of work we're delegating to it? Because if you can answer that precisely, you'll know whether you're using the tool brilliantly or putting your business on a foundation of pattern matching where it actually needs reasoning.

For years, AI remained conversational — you asked it something, it answered. The relationship was transactional. The humans were in charge. The AI was the tool.

But something has shifted. AI agents are now autonomous workers. They're navigating your cloud infrastructure, diagnosing failures, executing repair sequences. They're moving through web interfaces like humans do, finding information, filling forms, completing tasks. They're doing work without being asked step-by-step what to do next. They're doing work that looks like work.

Microsoft researchers released a framework called AgentRx that addresses an urgent problem: when these autonomous agents fail, we don't actually understand why. You have an incident. The agent was supposed to fix it. The agent failed. Now what? In human teams, you'd debrief — what went wrong, why, what did we learn? With AI agents, you mostly get silence. A failed action. No clear reason.

This matters because autonomous agents are becoming part of operational work in serious ways. They're managing systems that keep businesses running. When they fail, the consequences cascade. And right now, we don't have systematic methods to understand the failure.

When AI agents start doing autonomous work, they need to be as accountable as human colleagues — or riskier.

AgentRx proposes something that sounds simple but is actually complex: treat the agent's reasoning like you'd treat a worker's reasoning. When something goes wrong, trace back through the decisions. What did the agent understand about the situation? What goals did it think it was pursuing? Where did its reasoning diverge from what should have happened?

This is harder than it sounds because agents don't think like humans. Their internal states don't translate cleanly to human explanation. But the researchers argue that developing systematic debugging approaches is essential — not because it's good science, but because it's necessary safety practice.

Consider a concrete case: an agent is managing cloud infrastructure. It detects an anomaly, diagnoses a failing service, decides to restart it. The restart cascades into other systems and causes a larger failure. Now what? A human operator, confronted with this result, would be able to explain their reasoning: I saw this pattern before, it usually indicates this problem, restart usually fixes it. The agent can't give you that narrative. It's just a sequence of actions and a bad outcome.

AgentRx proposes debugging frameworks that let you systematically examine: Did the agent correctly perceive the situation? Did it use the right contextual information? Were its goals aligned with what you actually need? Where did the reasoning break down?

This is the unglamorous work of building AI systems that can be part of operational teams. It's not about making agents smarter. It's about making them transparent enough to hold accountable. And accountability is what allows you to actually trust them with real work.

The deeper point is this: AI agents are becoming colleagues in the sense that matters most — they're doing autonomous work that affects outcomes. And colleagues need to be understandable. You need to know why they made the decisions they made. You need to be able to learn from their failures. You need to be able to hold them accountable when something goes wrong.

Right now, most AI agents are opaque in exactly these ways. AgentRx is a step toward changing that. Not toward making agents perfect — they won't be. But toward making them legible. Toward building the kind of understanding between humans and AI systems that real collaboration requires.

Which raises a question that should matter more than it currently does: if you're deploying autonomous AI agents into critical work, what's your systematic method for understanding when they fail? Because "we don't really know" isn't a strategy. It's a liability.

There's something deeply human about expertise. An experienced surgeon doesn't carry every operation she's ever performed in her head. But she knows something — a feel for tissue, a sense of when something's wrong, an intuition that guides her hands. She has distilled years of experience into something she couldn't fully articulate if asked. It's knowledge, but it's not information.

This is what Microsoft researchers are trying to solve for AI agents. The problem sounds strange: they gave agents more memory and the agents got worse. They accumulated logs of every interaction. The logs grew enormous. The agents had to search through irrelevant detail to find what mattered. The signal got buried in noise.

So they built something called PlugMem. Instead of storing raw interactions — everything the agent did — PlugMem transforms experience into reusable knowledge. It distils patterns. It abstracts lessons. It creates something closer to what humans do when they develop expertise.

The parallel is striking. When you're learning to do something — to code, to negotiate, to manage people — you start with raw experience. You do the thing. You fail sometimes. You adjust. But you don't just accumulate failures. Over time, you begin to extract principles. Patterns you've seen before. Rules of thumb that work. You develop something you might call intuition but which is actually distilled knowledge.

Memory becomes useful only when it's been transformed into knowledge.

Expert humans are economical with what they remember. Not because they've forgotten, but because they've abstracted. A great chef doesn't remember every dish she's made. She's extracted principles about heat, flavour, timing. A experienced manager doesn't remember every difficult conversation. She's abstracted what works with different kinds of people, different contexts. Experience has been transformed into knowledge.

This is what PlugMem does. It watches the agent work. It extracts the patterns. It creates something like a condensed version of experience that the agent can actually use. Not a log of raw events — a distilled understanding.

The work implications are significant. As organisations deploy AI agents to handle ongoing tasks — customer service, technical support, operational management — these agents accumulate experience. The question is what they do with it. Without something like PlugMem, they accumulate noise. With it, they accumulate understanding.

And understanding is what allows an agent to handle novel situations better. When a customer service agent encounters a situation it hasn't seen, but has understood the principles underlying customer interactions, it can reason through. When a technical support agent faces an edge case, but has extracted principles about system behaviour, it can diagnose. The abstraction from experience is what makes experience actually useful.

This mirrors something important about human work. The difference between someone who's been doing a job for two years and someone who's been doing it for twenty isn't that the veteran remembers more. It's that the veteran has transformed memory into knowledge. Has extracted principles. Has developed intuition that's actually distilled expertise.

Which suggests something about how we should think about deploying AI agents into real work. They shouldn't accumulate experience like filing cabinets. They should transform experience into knowledge. And that transformation — that distillation of what matters — is where expertise actually lives.

The researchers aren't claiming PlugMem is perfect. But they're addressing something fundamental: the difference between having a memory and knowing how. And that difference is the difference between a system that processes information and a colleague that develops competence.

There's a particular kind of problem in AI deployment: the benchmark works perfectly, but when you roll the system into real work, it fails at exactly the thing you needed it to do.

Why? Because benchmarks measure performance on tasks. They don't explain what's actually driving success or failure. A system might ace a test without having developed the capability the task required. It might have found a pattern correlation. It might have memorised. It might have learned something domain-specific that doesn't transfer.

Researchers at Microsoft have developed something called ADeLe that addresses this gap. Instead of just reporting that an AI system performs at 85% accuracy on a task, ADeLe asks: what underlying capabilities is the system actually using? Where does it fail systematically? What would predict success on tasks it hasn't seen?

The implications are practical and significant. If you understand what capabilities an AI system actually has — not just how well it performs on benchmarks — you can design work processes that play to those strengths. More importantly, you can identify which parts of a job require human judgment that AI can't replicate, and protect those.

Understanding AI's actual capability profile is more useful than knowing its benchmark scores.

The research identifies several patterns. There are capabilities AI is genuinely strong at. Recognising visual patterns in images. Understanding text semantics when context is rich. Predicting outcomes when historical patterns are stable. When you have tasks that require these specific capabilities, AI is often transformatively useful. You can reduce cognitive load and let people focus on judgment.

But AI consistently struggles with others. Reasoning about what will happen in genuinely novel circumstances. Understanding causation versus correlation. Recognising when a situation falls outside the patterns it was trained on — exactly the moment when human judgment becomes most valuable.

The insight is this: it's not that some tasks are AI-suitable and others aren't. It's that tasks break down into constituent capabilities. And when you understand which capabilities matter for which part of the task, you can distribute the work rationally.

A loan approval process requires credit risk assessment (AI is good at this — it's pattern matching against historical lending data) and character assessment (AI is poor at this — it involves reasoning about novel circumstances and human intent). So you design the system to use AI for the quantifiable part and protect the role of human judgment on the part that requires it.

A diagnosis process requires pattern matching against medical literature (AI strong) and contextual reasoning about unusual presentations (AI weak). Protect the diagnostic judgment. Use AI for the knowledge-lookup part.

This is a different conversation than most organisations are having. Most are asking: can we automate this job? Can we do this with less labour? ADeLe suggests a different question: which parts of this job involve capabilities AI has, and which involve capabilities it doesn't? How do we distribute the work to play to actual strengths?

The researchers aren't claiming they've cracked the problem. ADeLe still involves making judgments about what counts as a "capability." But they're shifting the conversation from: does it work on the benchmark? To: what is it actually good at? And that shift changes everything about how you deploy it.

It means you're less likely to automation-first design. You're more likely to thoughtful augmentation design. You use AI where it's genuinely strong, which often means you can accelerate the parts of work that are highest-volume, lowest-judgment. And you protect the parts that require human reasoning.

Which suggests that the organisations building the most durable human-AI work arrangements aren't the ones pushing automation hardest. They're the ones who've developed precision about what their AI systems are actually capable of.

Imagine a robot tasked with cleaning a kitchen. Simple work. Routine. The kind of thing that seems like it should be easiest to automate. But watch what happens.

The robot observes the space. It identifies objects — a plate on the counter, a bottle on the shelf, crumbs on the floor. It plans a sequence of actions. It reaches for a cloth. And then something unexpected. The cloth is damp. The bottle isn't quite where it seemed. The plate moves when the robot tries to push past it. Small things. But they break the plan.

This is what AsgardBench, a new benchmark from Microsoft researchers, reveals. Robots can now observe environments and plan actions at a level that would have seemed miraculous a few years ago. The progress is genuine. But the gap between planning in ideal conditions and adapting when conditions break is where the real challenge lives.

And it's a gap that reveals something important about human work.

Physical work — the work that happens in spaces, with objects, with other people — requires constant adaptation. A care worker doesn't just execute a care plan. They respond to the resident's mood, the weather, what's available. A tradesperson doesn't just follow a procedure. They adapt when they reach the expected spot and find something different. A kitchen manager doesn't just execute a menu. They adapt when a delivery doesn't arrive or a customer has an allergy.

This adaptation isn't a rare edge case. It's the constant texture of physical work. Things never go exactly as planned. The world is too complex, too particular, too contextual.

The hardest human work isn't following a plan. It's knowing what to do when the plan breaks.

What the researchers document is that robots can now get quite far following a plan. They can observe, they can reason about sequence, they can adjust for modest variation. But the moment something genuinely unexpected happens — something outside the pattern they've learned — the competence drops sharply. They don't know what to do because they haven't learned what a solution looks like.

Humans do something different. They don't just follow the plan. They hold the purpose of the work in mind. When the plan breaks, they don't stop. They reason toward the goal from where they actually are. A robot reaches for a cloth. The cloth is wet. The robot struggles. A human reaches for a cloth, finds it wet, thinks: what's the goal here — to clean. What can I use instead? Reaches for a paper towel. Continuous. Fluid. Purposeful.

This matters enormously for thinking about which human work will remain distinct and valuable. It's not the routine parts. Those will increasingly be automatable. It's the parts that require reasoning when the plan breaks. The parts that require holding purpose in mind and navigating toward it despite unexpected obstacles.

In healthcare, it's not the routine care. It's the patient whose symptoms don't fit the diagnosis. In manufacturing, it's not the standard process. It's the moment when something goes wrong and you have to reason your way to a fix. In service work, it's not the standard interaction. It's the customer with an unusual need and you have to figure out how to help them.

The researchers aren't arguing that robots shouldn't improve. They should. But they're documenting that a crucial gap remains. And that gap corresponds almost exactly to the human work that's hardest to automate because it requires reasoning, presence, adaptability.

Which suggests something about how to think about automation and work futures. The automation that's most valuable isn't replacing human workers. It's removing the routine from their work. Teaching the robot to clean the obvious mess so humans can focus on the situation that doesn't fit the pattern. Automating the standard diagnosis so doctors can focus on the patient who doesn't fit. Creating space for the judgment, the presence, the adaptive reasoning that makes work distinctly human.

The benchmark shows we're making progress on planning. But it also shows we haven't solved the harder problem: what it means to adapt. And until we do, human presence in physical work — the capacity to respond to what actually happens — remains unreplaceable.

There's a particular kind of work intelligence that isn't about being clever in the moment. It's about holding a goal in mind across a long sequence of actions, keeping the purpose clear even when you're deep in detail, knowing when to adapt and when to hold the line.

A craftsperson working on a complex project. A researcher building an experiment. A project manager steering toward a distant deadline. A mentor developing someone's capacity over time. These people aren't solving problems in isolation. They're navigating toward goals that might be weeks, months, or years away, making decisions at each point that either bring them closer or pull them away from the destination.

This is long-horizon reasoning. And Microsoft researchers have now documented, precisely, how far AI remains from mastering it.

GroundedPlanBench is a benchmark for robot manipulation planning. It tests whether a robot can use vision and language understanding to plan a sequence of actions over multiple steps. But here's what makes it interesting: the sequences are long. The goal is distant. The path requires decisions that only make sense if you're holding the final goal in mind.

Imagine a robot tasked with arranging a kitchen. It needs to move items to shelves, organise by type, create space for new items, arrange for accessibility. That's not five steps. That's thirty steps. And every step decision cascades. Move this first and you have space for that. Move that first and you've blocked yourself from something else.

Humans do this routinely. A chef organising a kitchen. A parent managing a household. A teacher planning a curriculum. We hold the long-term goal, understand the constraints, and navigate the sequence. We can do it because we understand the principle — what matters and why — and we're not just following a procedure. We're reasoning toward purpose.

The work that requires holding a goal across dozens of steps without losing the thread remains stubbornly, distinctly human.

The research documents that vision-language models — the kind of AI that can see images and understand language — still struggle with this. They can make good individual decisions. But across a long sequence, they lose the thread. They become myopic. They optimise for the next step without considering whether it serves the long-term goal.

This is revealing because long-horizon reasoning is everywhere in skilled work. It's not a rare edge case. A surgeon planning a complex procedure. A mechanic diagnosing a problem that will require multiple repairs in sequence. A manager navigating an organisation through change. An architect designing a space that will serve multiple purposes. All of these require holding a complex goal across many steps.

And all of them remain difficult for AI precisely because AI is fundamentally myopic. It's built on pattern matching within defined contexts. Long-horizon reasoning requires something different — it requires abstract purpose, the capacity to evaluate whether you're getting closer to a goal you won't reach today, the judgment to sacrifice short-term efficiency for long-term progress.

The researchers aren't claiming this is unsolvable. But they're documenting that the gap is significant. And they're suggesting that this gap corresponds exactly to the skilled work that remains most distinctly human.

This has implications for how organisations think about human-AI collaboration. The work you're most likely to be able to automate is the work that's episodic — solve this problem, answer this question, categorise this item. The work you're least likely to be able to automate is the work that's diachronic — hold this goal, navigate this long sequence, keep this vision alive across months of detailed decisions.

Which means the work you should be protecting, the expertise you should be developing, is the work of long-horizon thinking. The capacity to hold a distant goal and navigate toward it. To see how individual steps serve larger purpose. To know when to stay the course and when to adapt.

This is the work of strategy, stewardship, mentorship, craft. It requires presence over time. It requires understanding principles, not just patterns. It requires the kind of judgment that doesn't emerge from data but from lived experience of pursuing something difficult over time.

Ironically, in a moment when everything is optimising for the short horizon — quarterly targets, immediate metrics, fast feedback loops — the most valuable human work is the capacity to hold the long horizon. To think beyond the next step. To navigate toward distant goals despite obstacles and temptations to optimise locally.

Which raises a question that should matter more to leaders and organisations: are you building capacity for long-horizon thinking? Or are you inadvertently optimising it away in pursuit of measurable short-term progress?

For years, the most powerful AI has been available only to the largest organisations. The models were big. They cost millions to train. They required serious compute to run. So they belonged to tech giants. Everyone else got the sanitised consumer versions — helpful but limited, designed to be safe for anyone to use, which also meant designed to be useful for almost no one.

But something has shifted. Open-weight models are changing that. These are models that organisations release not just as APIs but as actual weights — the underlying data that makes the model work. You can download them. Run them on your own infrastructure. Modify them. Fine-tune them for your specific work.

Microsoft's recent release of Phi-4-reasoning-vision-15B is significant for one reason: it's a 15-billion-parameter multimodal reasoning model. It can see. It can reason about what it sees. It can explain its reasoning. And it's open-weight.

This is early days, but the implications are starting to become clear. A small manufacturing company can now download this model and train it on their specific equipment. When something breaks, the model can see what's wrong, reason about the cause, suggest repairs. No dependence on distant vendors. No paying per API call. No waiting for features from big tech.

A building maintenance team can deploy it to inspect for problems. A healthcare clinic in a place where specialist diagnostics are hours away can run it locally on medical images. A construction team can use it to quality-check work before it's accepted. A small farm can use it to scout for crop problems.

The capacity to see and reason about physical work is moving from the cloud to the edge, from corporations to communities.

The first-mover advantage that big organisations had — access to the best AI — is eroding. And that changes the game.

There's something important about the fact that this is multimodal. That it doesn't just read text but understands images. Because so much skilled work is about seeing. A mechanic diagnosing a broken engine sees the problem. A nurse assessing a patient sees the condition. A farmer checking soil sees what needs attention. This has always required human presence. Presence and vision working together.

Now there's an alternative. The seeing can be distributed. A piece of equipment can have a camera and run this model locally. A diagnostic can happen instantly, at the site, in the language of the people doing the work. The reasoning can be explained. It can be contested. It can be learned from.

This is different from previous automation waves. Previous waves concentrated capability and often concentrated power along with it. You needed to go to the factory. You needed access to the system. You needed permission from the keeper of the technology. Open-weight models decentralise this. The technology can be deployed where the work is.

There are risks, obviously. A poorly trained model giving dangerous advice. Models trained on biased data causing harm. The democratisation of capability also democratises the capacity to do things badly. But the alternative — keeping powerful tools locked behind corporate gateways — hasn't been working either. It's kept capability from the people who could use it most.

The real story here isn't about Phi-4. It's about what happens when you release powerful tools and trust communities to use them. What happens when a small team in a place with limited resources can download the same reasoning capacity that big tech companies use. What kinds of work become possible? What kinds of problems get solved?

This is happening now. The research community is moving fast. The models are getting better. And they're becoming available to anyone willing to run them.

Which means the question for the next phase of work isn't: will AI change work? That's settled. The question is: who gets to decide how it changes? Will it remain concentrated in big organisations, or will it flow out to the communities doing the actual work?

The release of Phi-4-reasoning-vision suggests the answer is starting to shift. And that shift matters more than any individual improvement in model performance.

There's something unusual about what Microsoft Research has done with their new podcast, "The Shape of Things to Come." They've started with a premise that's almost radical in a moment of rapid AI deployment: let's slow down. Let's ask fundamental questions. Let's think carefully about what we actually know and what we're assuming.

The podcast, led by Doug Burger, explores basic questions: What is intelligence? How does it emerge? What are the actual limits of current AI approaches? What would genuinely new capabilities look like? These aren't questions with clear answers. They're questions that require sustained, careful thinking.

And in a moment when the conversation around AI is driven by pace — move fast, deploy faster, stay ahead of the curve — this feels almost countercultural.

The work of rigorous thinking is work. It's labour. It takes time. It requires sitting with uncertainty rather than rushing to conclusions. It requires examining evidence rather than accepting narratives. It requires intellectual humility — the acknowledgement that most of what's happening with AI is still poorly understood.

This matters because the conversation around work and AI has become increasingly frantic. Every announcement about new capability triggers predictions about job displacement. Every concern about misuse triggers dismissal as scaremongering. The discourse has become binary — utopian or dystopian, with very little space for the actual, complex, uncertain middle.

In a moment of rapid change, the capacity to think carefully and publicly is increasingly valuable.

The podcast doesn't solve this. But it models something important: what rigorous thinking looks like. Take a complex question. Don't rush. Listen to researchers who've spent years studying the problem. Consider multiple perspectives. Acknowledge what you don't know.

This is the opposite of the speed culture that dominates most tech discourse. It's the opposite of the demand for immediate takes on every development. It's the opposite of the reduction of complex questions into simple narratives.

And it's particularly important in the domain of work, because decisions about how to deploy AI at work are being made now, by people with incomplete understanding, on compressed timescales, with high stakes for the people affected.

A manager decides whether to deploy an AI system to monitor worker productivity. A training officer decides how to allocate scarce resources between automation and developing human capability. A company decides what skills to invest in for its workforce. These are decisions that require more than speed. They require understanding.

And understanding requires the kind of sustained, careful thinking that "The Shape of Things to Come" is attempting to model.

What's striking about the research community's willingness to do this is that it's not fashionable. It doesn't move fast. It doesn't promise solutions. It asks questions, carefully, in front of an audience, with the understanding that the audience might challenge the thinking or contribute to it.

This is different from either hype or denial. It's the work of genuine intellectual engagement.

The podcast is also notably public. It's not a research paper in an academic journal that three people will read. It's accessible. It's conversational. It's an attempt to bring the actual questions researchers are asking into public view, rather than leaving the narrative to people selling products or people predicting doom.

Because the stakes are real. We're making decisions about work systems, about what kinds of jobs to automate, about what human skills to protect or develop. These decisions deserve more than speed. They deserve rigorous thinking.

Which raises a question that might matter most: in a moment of rapid change, where is the space for careful thinking? Who's creating it? Who's participating in it? And how do we make sure that the people making decisions about work and AI actually have access to rigorous understanding, not just fast takes?

This podcast is a small attempt at that. And small attempts at rigorous public thinking might be among the most important work being done right now.