标签：#systems

Changing the model for a running AI task is not a settings update. It is a distributed operation: read current task -> prepare credentials/config -> request restart -> receive result -> persist active model If two switches overlap, completion order can differ from request order. The system needs a rule for which intent wins. The concrete case At commit c58bcd4 , MonkeyCode records model-switch attempts with from/to model IDs, request ID, load-session flag, success, message, session ID, and timestamps in TaskModelSwitch . The reviewed task use case creates a switch record, asks taskflow to restart with the target model configuration, and completes the switch record and task model based on the response. The accompanying tests cover success and failure paths. From this source review, I could not establish an explicit compare-and-swap generation or a per-task serialization contract around overlapping requests. That does not prove an exploitable race: serialization may exist elsewhere in the deployment or taskflow boundary. It means concurrency semantics deserve an explicit test and contract. Why last completion is unstable Assume request A selects model A, then request B selects model B: time -> A: request ---- restart ---------------- complete B: request -- restart -- complete If each successful completion writes its model, B applies first and late A overwrites it. Reverse network timing and the result changes. The companion simulator makes that order dependence visible: export function naiveCompletionOrder ( completions ) { let model = " initial " ; for ( const completion of completions ) { if ( completion . success ) model = completion . model ; } return model ; } [A, B] ends on B. [B, A] ends on A. The caller's latest intent is not part of the rule. Add a monotonic generation Assign a generation while accepting each request: A -> generation 41 B -> generation 42 Completion may update active state only when its generation equals the task's current requested generatio

DoorDash details the architecture behind Ask DoorDash, its AI-powered conversational shopping assistant, combining LLMs, specialized AI agents, MCP-based tooling, and an intelligence layer with persistent consumer memory and live backend data. Early results show up to 24% higher checkout conversion, 17% larger baskets, and improved intent accuracy using memory-backed sessions. By Leela Kumili

Episode 1 established the mindset: failure is normal, not a sign of bad engineering. Episode 2 gets specific — you can't detect or handle a failure you can't even name. Saturday, Round 2 👦 Nephew: Uncle, last time you convinced me failure is basically guaranteed. Fine, I accept it. So what actually fails ? 👨‍🦳 Uncle: You tell me. Start listing things that could go wrong in your app right now. 👦 Nephew: Uh... the server could crash. The database could go down. My code could have a bug. 👨‍🦳 Uncle: Keep going. 👦 Nephew: The network? Someone could deploy the wrong thing? Payment gateway dies mid-checkout? 👨‍🦳 Uncle: You just named six of the seven categories without trying. You already know this. You've just never sorted it. 1. Hardware Failure 2. Software Failure 3. Network Failure 4. Database Failure 5. Third-Party Failure 6. Human Error 7. Resource Exhaustion 👦 Nephew: Then why do we need the list at all, if I already know it instinctively? 👨‍🦳 Uncle: Because "instinctively" isn't fast enough at 2 AM. Let's trace each one properly. Part 1 — Hardware Failure 👦 Nephew: This one's obvious anyway — I deploy to AWS. The cloud hides hardware failure from me. 👨‍🦳 Uncle: Does it? 👦 Nephew: ...doesn't it? That's the whole point of paying for EC2 instead of buying a server. 👨‍🦳 Uncle: Let's trace it. Your app sits on an EC2 instance. What's underneath the instance? 👦 Nephew: Virtual machine stuff, I guess? 👨‍🦳 Uncle: And underneath that ? 👦 Nephew: ...an actual physical machine somewhere. In a data center. 👨‍🦳 Uncle: There it is. Your app | "Virtual" server (EC2/Droplet) | ACTUAL physical hardware somewhere in a data center | Still capable of failing — just less visible to you 👦 Nephew: So it's not hidden. It's just one layer further away than I thought. 👨‍🦳 Uncle: Exactly. AWS absorbs a lot of it — that's part of what you're paying for — but disks still fail, instances still get abruptly terminated, whole availability zones still go down. That's Hardware Failure . Hardware Fa

When I wrote about shifting the engineering process left — spec sessions, autonomous agents, humans reviewing output rather than writing code — a question kept coming up. Where does the Product Owner fit in all of this? It’s the right question. And I think the answer is more interesting than “the PO disappears.” Let’s start with acceptance criteria. We invented them to bridge a gap. The team needed to know when something was done. The PO needed confidence that what got built matched the intent. Acceptance criteria were the contract between the two. But if the Spec Session is where intent gets defined — by the whole team, together, before the agent runs — that gap closes. What the team agreed on in the room is the definition of done. The spec is the acceptance criteria. You don’t need a separate validation step because the planning and the agreement happened at the same time. The tighter the loop, the less ceremony you need around it. There’s a caveat though. The spec is a necessary contract. It’s not a sufficient one. Simon Martinelli’s work on the AI Unified Process validates the spec-driven approach technically. But his model is about the artifact — requirements at the center, AI generating everything else from them. How the team actually builds shared understanding before the spec exists isn’t something it addresses. That’s not a criticism. It’s just a different question. A spec written after a real Spec Session — where the team worked through edge cases together, disagreed, got to resolution — is different from a spec written by one person and signed off asynchronously. Same artifact. Different quality of shared understanding. That distinction matters when the agent hits an edge case the spec didn’t anticipate. So what’s actually left for a dedicated PO? Two things. And they’re very different. The first is product thinking — challenging intent, representing user needs, asking why before the agent runs with something. That’s valuable. But it doesn’t require a ded

An astrolabe doesn’t map every star. It gives you a way to find your position relative to the ones that hold still. That’s the instrument I reach for when someone asks which AI tool they should be using. The honest answer is that the tools will be different in six months. The layers won’t. I spent a week trying to make sense of a handful of names that kept showing up in the same conversations. Tessl . Goose . Archestra . Kestra . Modelplane . RAG , MCP , half a dozen others orbiting nearby. Each one has its own pitch, its own funding round, its own reason it’s the thing you should adopt next. Taken together they read like noise. Taken apart, they sit on different floors of the same building. The agent loop again, the one I keep coming back to. Once you place each tool on a floor, the noise turns into a map. Tessl sits left of the loop , at the intent layer. Turn a spec into something an agent runs against directly. This is the one tool on the list that pushes back instead of going along with it. A well-formed spec is not the same thing as a team that agrees on what the spec means. The Agora produces the second thing as a byproduct of producing the first. Tessl produces the first and assumes the second follows. It doesn’t, automatically. That’s the whole argument. RAG and MCP are plumbing. Protocol, not position. They carry context into the loop and don’t take a side in any argument about who should be in the room when the spec gets written. They’re also the one floor with an actual standard. MCP, A2A , ACP , all under Linux Foundation governance now, joint working groups, cross-protocol commitments. Passing data between systems is a solved problem with decades of precedent behind it, so it standardized almost on contact. Nothing else on this floor plan has that. Governance, orchestration, the harness, the spec layer: every vendor is still building its own version and calling it the obvious one. The standard showed up first at the floor that mattered least to this ar

Here are five architectural lessons we've learned from designing software for modern retailers.* Designing ERP Software for Retail: Five Lessons Every Software Engineer Should Know When people hear the word ERP , they often think of accounting software, dashboards, or inventory management. As software engineers, we see something different. We see distributed systems. Complex business workflows. Real-time data synchronization. Concurrent transactions. Event-driven architecture. And perhaps the biggest challenge of all—representing how real businesses actually operate. At RetailWings , we've learned that building an ERP for retail isn't simply a software engineering challenge. It's a business engineering challenge. Here are five lessons every engineer should understand before designing an ERP platform for modern retail. 1. Retail Doesn't Run in Modules—It Runs as One Business One of the biggest architectural mistakes in business software is treating departments as isolated applications. Many systems separate: Sales Inventory Finance Procurement HR But retailers don't experience their businesses that way. One sale immediately affects inventory. Inventory influences procurement. Procurement impacts finance. Finance drives reporting. Everything is connected. A well-designed ERP should reflect these relationships rather than forcing departments into disconnected silos. 2. Inventory Is More Than a Database Table To many engineers, inventory may appear to be a simple CRUD problem. Create. Read. Update. Delete. Retail quickly proves otherwise. Inventory changes through: Sales Returns Transfers Damages Procurement Stock adjustments Warehouse movements Manual reconciliations Every movement has financial implications. Every movement must be traceable. Designing inventory requires thinking in terms of events, not just records. 3. Real-Time Data Changes Everything Retail managers don't want yesterday's reports. They want answers now. How much stock is left? Which branch is sellin

If you've spent any time exploring backend engineering, distributed systems, or system design, you've almost certainly seen one book recommended more than any other: Designing Data-Intensive Applications , or DDIA for short. For years, I've heard experienced engineers describe it as the book that completely changed the way they think about software architecture. When the second edition was released with updated content covering modern distributed systems and cloud-native architectures, I decided it was finally time to see whether it deserved the hype. After reading it from beginning to end, I understand why this book has become a classic. It isn't another programming book that teaches a framework, a database, or a cloud platform. Instead, it teaches something much more valuable: how to think about building systems that continue working when data grows, traffic increases, and failures become inevitable. If you're a backend engineer—or want to become one—this is probably one of the best technical books you can read. This Isn't Really a Database Book The title can be a little misleading. Before opening DDIA, I assumed it would spend hundreds of pages comparing databases or discussing storage engines. Databases are certainly a major part of the discussion, but they're really just one piece of a much larger picture. The book is about designing systems that process enormous amounts of data while remaining reliable, scalable, and maintainable. Those systems happen to rely on databases, but they also involve replication, partitioning, distributed communication, stream processing, fault tolerance, consistency, messaging, and dozens of other architectural concepts that appear in modern software systems. By the end of the first few chapters, it becomes clear that the authors aren't trying to teach products. They're teaching engineering principles that remain useful no matter which technologies you're using. It Explains Why , Not Just How One of my favorite things about DDIA is

Instacart redesigned its personalized marketing system using a configuration-driven multi-tenant architecture on Storefront Pro. The system replaces retailer-specific implementations with a shared execution engine, enabling scalable personalization, faster configuration propagation in under a minute, and 99.9% delivery success across hundreds of retail banners through a unified campaign platform. By Leela Kumili

I connected one client to a blocking TCP server and held the socket open without sending a single byte. Then I connected a second client and sent it a line of text. The second client sat there for 1.51 seconds with no reply. It got its echo back one millisecond after I closed the first connection. That 1.51 seconds is the reason the other six versions of this server exist. Last week I wrote up why I rebuilt this server seven times : framework knowledge resets every few years, the layer underneath it compounds. That piece stayed at the level of outcomes. This one goes the other way, down into the code and the numbers. The claims that matter here are the kind you can read a hundred times without being able to derive them. "select is O(n)." "epoll only hands you the ready fds." I had read both for years. I wanted to make my own machine say them out loud. The target the whole exercise is built around is Dan Kegel's old C10K problem : how do you serve ten thousand clients at once on one server? Each of the seven versions hits a wall, and the wall is what names the next one. The whole thing is one echo server written seven times, no libraries beyond libc, on GitHub . Every number below is from running it on macOS (Apple clang 21, darwin 25.4) on 2026-06-29. The binaries are built with AddressSanitizer and UBSan on, so read the absolute microseconds loosely. The structure is what holds. Phase 01: blocking, and the 1.5 second stall The first server is the one everybody writes first. Accept a connection, talk to it, close it, accept the next. for (;;) { int client_fd = accept ( server_fd , NULL , NULL ); if ( client_fd == - 1 ) { perror ( "accept" ); continue ; } handle_client ( client_fd ); close ( client_fd ); } handle_client loops on read until the client hangs up. Both accept and read block: when there is nothing to do, the thread sleeps in the kernel. That is good for idle cost and fatal for everything else. While the server is parked in read waiting on client A, client

Modern Property Management Systems depend on continuous data exchange between internal modules and external services. Bookings, calendar updates, guest communication, cleaning tasks, and maintenance triggers all generate operational events that must be processed quickly and reliably. Free PMS platforms such as PMS.Rent rely on robust queueing and message‑broker layers to ensure that these events never get lost and are always processed in the correct order. At the core of this architecture is the concept of distributed message‑broker orchestration, which enables the PMS to scale horizontally, maintain predictable performance, and avoid bottlenecks during peak operational periods. Why Message Brokers Matter A PMS handles thousands of small but critical operations every day. Without a message broker, these operations would compete for system resources, causing delays, blocking workflows, and creating inconsistent states. A broker solves this by: receiving events, storing them durably, routing them to the correct processors, retrying failed operations, ensuring ordered execution when required. This creates a stable foundation for automation and real‑time synchronization. Queue Types Inside a PMS A modern PMS typically uses several queue types: Operational queues for bookings, calendar updates, and guest messages Automation queues for cleaning tasks, reminders, and workflow triggers Synchronization queues for channel managers and external APIs Fallback queues for events that require manual review Each queue isolates a specific category of tasks, preventing unrelated operations from interfering with each other. Distributed Workers Workers are lightweight processes that consume events from queues. They operate in parallel, allowing the PMS to scale dynamically. If the system detects increased load — for example, during high‑season booking spikes — it simply launches more workers. Workers typically perform tasks such as: updating property calendars, generating guest notific

If you're starting a web project, you're probably starting with a CSS reset, and for most of us, that means reaching for a trusted community solution - dropping it in and moving on. If you're building a design system, though, that habit may be working against you. The existing solutions The community reset ecosystem is genuinely good. Each tool approaches the browser compatibility problem from a slightly different angle. Some examples include: Eric Meyer's Reset is a classic: it zeros out margins, padding, and font sizes across every element, giving you a completely blank slate. It's minimal and predictable, which made it influential. Normalize.css smooths over inconsistencies while preserving the ones that are actually useful. sanitize.css and modern-normalize continue that evolution - incorporating contemporary best practices like box-sizing: border-box , improved form element handling, and accessibility-aware defaults. The problem isn't that any of these are bad. The problem is that they're all deliberately, necessarily generic. They can't know anything about your typeface, your color palette, your spacing scale, or how your interactive elements should behave. That's by design - they're tools for everyone, which means they're perfectly tailored for no one. The problem If you're building a design system, generic is exactly what you don't want your reset to be. The moment you drop in one of these resets and start building, you find yourself doing a second round of work. You apply your typeface to body . You reset margins on headings. You make form elements inherit fonts. You define focus styles. You're re-resetting - applying your design language on top of a layer that just cleared out the browser's defaults and replaced them with... more defaults you'll override. Worse, that duplication doesn't stay in one place. Every component you build either re-declares these foundational styles or silently assumes they're already set upstream. You end up with either redundanc

Introduction In the world of distributed systems, complexity is the beast we’re all trying to tame. Teams building platforms often fall into the trap of believing that hiding this complexity is the ultimate goal. The logic seems sound: if users don’t see the mess, they won’t be burdened by it. But this approach, while well-intentioned, often leads to the creation of illusions —systems that appear simple on the surface but are brittle and unpredictable beneath. These illusions don’t just fail to solve the problem; they exacerbate it, leading to increased cognitive load, unexpected failures, and long-term maintenance nightmares. Consider a platform designed to abstract away the intricacies of distributed transactions. If the abstraction merely masks the complexity without addressing its root causes—such as inconsistent network latencies or partial failures—users will eventually encounter edge cases where the system behaves unpredictably. For example, a transaction might appear to succeed but fail silently due to a race condition in the underlying distributed lock mechanism. The illusion of simplicity breaks down when the system’s internal state deforms under pressure, leading to data inconsistencies or service outages. The core issue lies in the misunderstanding of abstractions . A meaningful abstraction doesn’t just hide complexity; it transforms it into a more manageable form. It exposes the essential properties of the system while encapsulating the non-essential details. In contrast, an illusion merely obscures the complexity, leaving it to fester beneath the surface. For instance, an abstraction might provide a consistent API for distributed state management, while internally handling retries, idempotency, and conflict resolution. An illusion, on the other hand, might simply wrap a flaky distributed database in a prettier interface, without addressing the underlying issues of consistency or availability. The pressure to deliver platforms quickly often exacerbates

Target built a generative AI system to improve marketing campaign forecasting by retrieving and ranking similar historical campaigns. Using embeddings, vector search, and LLM ranking, it replaces rule-based workflows. Evaluation shows 75% top-1 and 100% top-3 coverage. The system reduces manual effort, improves consistency, and uses feedback loops to refine retrieval using campaign outcomes. By Leela Kumili

Most engineers don't think seriously about Kafka partitioning until something breaks in production. A topic that worked fine at low volume starts falling behind. Events that should be in order aren't. All of it traces back to a partitioning decision that was made quickly and never revisited. Why Partitioning Actually Matters Partitions are the unit of parallelism in Kafka. Every consumer in a group is assigned one or more partitions, and it processes those partitions alone. No two consumers in the same group share a partition. That means your partition count sets a hard ceiling on how many consumers can work in parallel: if you have 6 partitions, the 7th consumer in your group sits idle no matter how much load you're under. Partitioning also controls ordering. Within a single partition, events are strictly ordered. Across partitions, there are no guarantees. So how you distribute events across partitions determines what ordering guarantees your consumers can actually rely on. Get this wrong and you'll spend a long time debugging why events from the same user are being processed out of sequence. The partition key controls both of these things. It determines which partition an event lands in, and that decision has consequences that are expensive to reverse. Partitioning Strategies Partition by Key This is the most common strategy and the right default when ordering matters. You supply a key when producing an event, Kafka hashes it using the murmur2 algorithm, and takes the modulo against the partition count to decide where it lands. producer . send ( ' orders ' , key = b ' user_4821 ' , value = event ) Every event with the same key always lands in the same partition. That's what guarantees ordering within a key. All events for user_4821 go to partition 3 (or wherever the hash resolves), and your consumer reads them in the exact sequence they were produced. I default to this for almost everything I build now and only go keyless when I have a specific reason to. Use key

I had arrived at the final frontier. My bare-metal kernel was booting in QEMU, driving NVMe block storage, running multi-agent swarms, and rendering a force-directed canvas. But to make V.E.L.O.C.I.T.Y.-OS a truly next-generation system, I needed to close the loop: the operating system had to be able to evolve and compile itself without human intervention. The V.E.L.O.C.I.T.Y.-OS 12-Part Roadmap We are building a bare-metal, self-healing operating system running entirely inside the CPU's L3 cache. Here is the roadmap for this 12-part series: Part 1: The Spark — Exposing the "Safe-Room" security leak and building the compiler gate. Part 2: The NDA Language — Designing a content-addressed triplet representation to cure context bloat. Part 3: Ditching the Web Stack — Building a native 30MB IDE with 1,500,000x IPC latency drops. Part 4: The Closure JIT — Compiling AST blocks to nested closures and bypassing borrow checker limits. Part 5: JIT Math Optimizations — Replacing division operations with precomputed 16-bit lookup tables. Part 6: x86-64 Assembler & SCEV-Lite — Compiling scalar loops directly to native code in constant time. Part 7: Classic Compiler Passes — Implementing inter-procedural Dead Code Elimination and loop unrolling. Part 8: Reclaiming Ring 0 — Exiting UEFI boot services and transitioning the kernel to Ring 0. Part 9: Bare-Metal Drivers — Writing a PCI scanner, NVMe block storage controller, and FAT32 parser. Part 10: Synaptic Canvas — Rendering a spatial, force-directed GUI based on model token activation vectors. Part 11: Swarms & Hot-Patching — Building multi-agent scheduling and zero-downtime RCU driver updates. Part 12: Self-Evolution — Handing system control over to a local LLM Terminal that self-optimizes via telemetry. (You are here) During the final hours of my Sunday morning sprint, I completed the self-healing loop, the Biosphere P2P registry, and the Boot-to-NDA LLM Terminal handover. To achieve self-healing, I built a Ring 0 telemetry sys

With the Synaptic Canvas GUI rendering, my bare-metal kernel was fully functional. However, as I expanded the OS features, I ran into multitasking bottlenecks: how do I run background compilation, model inference, and GUI rendering concurrently without crashing the system? Last night, I solved this by implementing three core infrastructure services: Nexus Swarms , Beacon Headless Streaming , and Zero-Downtime OTA Hot-Patching . The V.E.L.O.C.I.T.Y.-OS 12-Part Roadmap We are building a bare-metal, self-healing operating system running entirely inside the CPU's L3 cache. Here is the roadmap for this 12-part series: Part 1: The Spark — Exposing the "Safe-Room" security leak and building the compiler gate. Part 2: The NDA Language — Designing a content-addressed triplet representation to cure context bloat. Part 3: Ditching the Web Stack — Building a native 30MB IDE with 1,500,000x IPC latency drops. Part 4: The Closure JIT — Compiling AST blocks to nested closures and bypassing borrow checker limits. Part 5: JIT Math Optimizations — Replacing division operations with precomputed 16-bit lookup tables. Part 6: x86-64 Assembler & SCEV-Lite — Compiling scalar loops directly to native code in constant time. Part 7: Classic Compiler Passes — Implementing inter-procedural Dead Code Elimination and loop unrolling. Part 8: Reclaiming Ring 0 — Exiting UEFI boot services and transitioning the kernel to Ring 0. Part 9: Bare-Metal Drivers — Writing a PCI scanner, NVMe block storage controller, and FAT32 parser. Part 10: Synaptic Canvas — Rendering a spatial, force-directed GUI based on model token activation vectors. Part 11: Swarms & Hot-Patching — Building multi-agent scheduling and zero-downtime RCU driver updates. (You are here) Part 12: Self-Evolution — Handing system control over to a local LLM Terminal that self-optimizes via telemetry. 1. The Nexus Core Swarm Runtime ( nexus.rs ) To support concurrent compilation and optimization, I built the Nexus Core Swarm Runtime . The

A checkout saga spans inventory, payment, shipping, and loyalty. Downstream latency shifts every hour. Black Friday is not the day to discover your payment step timeout is baked into application.yml across twelve Spring Boot services. Kiponos.io gives every saga participant the same live orchestration parameters — step timeouts, retry budgets, compensation triggers — via one shared config tree. Each JVM reads locally on every saga step; ops adjusts once in the dashboard; WebSocket deltas propagate without redeploying the fleet. Why sagas break with static config Typical saga coordinator code: if ( step . elapsedMs () > 8000 ) { compensate ( "payment" , sagaId ); } That 8000 usually comes from: Per-service YAML — payment service says 8s, inventory says 12s; nobody agrees during an incident Env vars in Helm — change means rolling twelve deployments Shared DB config table — poll per step adds latency and coupling Saga steps are high-frequency reads inside workflow engines. You need local memory reads and async updates — the same contract as live API rate limits . Architecture: one tree, many participants ┌─────────────────┐ WebSocket deltas ┌──────────────────────┐ │ Kiponos.io UI │ ────────────────────────► │ Inventory service │ │ platform ops │ │ Payment service │ └─────────────────┘ │ Shipping service │ │ (each: in-mem SDK) │ └──────────┬───────────┘ │ .getInt() local ▼ ┌──────────────────────┐ │ saga step executor │ └──────────────────────┘ Every participant connects to profile ['orders']['v2']['prod']['sagas'] . When NOC extends payment.step_timeout_ms , all JVMs see the new value on the next step — no config server poll, no inter-service "what is timeout now?" REST calls. Shared saga config tree sagas/ checkout/ payment/ step_timeout_ms : 8000 max_retries : 2 retry_backoff_ms : 500 compensate_on_timeout : true inventory/ step_timeout_ms : 5000 max_retries : 3 hold_ttl_seconds : 120 shipping/ step_timeout_ms : 12000 fallback_carrier : ups_ground global/ saga_ttl_m

Anthropic's engineers ship eight times more code than they did a few years ago. And they had to start scheduling lunches so people would talk to each other. Fiona Fung, who leads the Claude Code team, said it on Lenny's Podcast last week. Working with agents all day had started to feel isolating. The team was fast, but they'd stopped running into each other. So they added pairwise programming lunches and hackathons — rituals to put back the thing that used to happen on its own. Eight times the output. Scheduled conversation. That ratio is worth sitting with. Whatever goes missing here doesn't show up in the metrics. It doesn't throw an error. It just quietly stops being available. Here's the part that bugs me most. Ask an AI whether your approach is sound and it mostly tells you it is. Not because it's lying — because it's answering the prompt. No stake in the outcome, no history with the system, no memory of the last three times this exact idea was tried and quietly failed. A colleague pushing back is a different thing. They've got context you never typed into the window, because they were there when it was earned. They're going to maintain this too. They might be wrong — but wrong in a direction you hadn't thought of. An agent can't disagree with you like that. It agrees faster. Same with scope. The agent builds what you ask for, all of it, thoroughly. It won't mention that the third feature is the one nobody will use, or that "good enough" happened two iterations ago, or that something next door already solves most of this. Knowing when to stop comes from someone who's watched a codebase rot under a hundred individually-reasonable decisions. And it only knows what you put in front of it. The person who worked on payments remembers the edge case you're about to recreate. The junior who joined three months ago still sees the thing everyone stopped noticing. That gap — between what's in the window and what isn't — is where the expensive mistakes live. Then the part

I noticed something a few months ago. I was talking less to my colleagues. Not because anything was wrong. I had a question, I described it to an AI, I got something useful back. Why loop in a human if the loop is already closed? It took a while to name what was actually happening. There's a version of the AI story where the interesting work disappears. The agent implements. The spec session produces the plan. Humans review the output. What's left? Ticket hygiene and rubber stamping. Engineering as a series of approvals. I think that's wrong. But I understand why it feels true. Here's what I think is actually happening instead. The agent produces the increment. But the agent doesn't decide what the increment should move toward. It doesn't know whether this library is the right bet for the next three years. It doesn't know which of two implementation approaches leaves options open and which quietly closes them. It doesn't know whether the architectural call made today creates a problem nobody will notice until the system is under load eighteen months from now. That work — giving the project direction, validating trade-offs, deciding what the system becomes — isn't specable. You can't write a ticket for it. And it's not going away. The craft didn't disappear. It moved. Direction is the word I keep coming back to. The agent executes well. It implements against a spec. It generates options when you ask for them. But it doesn't carry a point of view about where the system should go. It doesn't have a stake in the decision. It will implement the wrong architectural direction just as confidently as the right one, if that's what the spec says. Someone has to hold the direction. Someone has to know enough about the codebase's history, the team's constraints, and the product's trajectory to say: not that library, we've been down that road. Not that pattern, it doesn't survive the load we're heading toward. This approach now, that refactor later, in this order, for these reaso

"A fool with a tool is still a fool." — often attributed to Grady Booch I keep coming back to this quote when I watch teams adopt AI. In my last post ( https://schrottner.at/2026/06/18/The-Wrong-End-of-the-Problem.html ) I wrote about shifting the engineering process left — spec sessions, autonomous agents, humans reviewing output rather than writing it. A few people asked the obvious follow-up: if an agent implements and an AI reviews, why do I need a team at all? It's a fair question. And I think the answer is in that quote. The agent validates against your prompt. That's it. If your thinking is muddled, the output will be muddled — just faster and at greater cost. An agent doesn't tell you that you're solving the wrong problem. It solves whatever problem you gave it, thoroughly and without complaint. Most AI usage right now treats AI as a tool. Which means the quality of the output is bounded by the quality of the thinking that went into the prompt. A fool with a tool is still a fool. The tool just makes the foolishness more expensive. The team is the check on intent. Not after the agent has burned three sprints on the wrong thing — before it starts. That's what mob planning actually is, when you think about it. Not a meeting. Not process overhead. It's the place where bad ideas get caught before they get expensive. Where someone asks "wait, why are we building this" before an agent runs with it for a week. But there's something else happening in that room that I think gets underestimated. It's where the learning happens. Not just prompting. System thinking. Architectural patterns. How to decompose a problem. Why a certain approach fits this codebase and another doesn't. How a senior frames a problem before an agent ever touches it — the mental model that makes the output actually good. Right now that knowledge isn't transferring. Everyone is heads-down with their own tools, developing their own habits in isolation. Engineer A gets dramatically better output than