标签：#agentic

Plus, a new deprecation policy ensures features aren't removed suddenly.

An Agent Harness is a comprehensive application layer that securely wraps a Large Language Model (LLM) to govern its memory, tools, execution boundaries, and deterministic policy enforcement. When engineers first transition from building simple conversational chatbots to fully autonomous AI agents, they typically make a critical mistake: they treat the Large Language Model (LLM) as the entire system. The reality is quite different. The LLM is not an agent. The LLM provides a reasoning engine, and nothing else. Everything else we build around that engine—the memory, the execution of tools, the planning capabilities, the routing of context, and the security boundaries—is the Agent Harness . Why an Agent Harness is Important If an LLM is the engine of a car, the harness represents the steering wheel, the brakes, the transmission, and the dashboard. When you give an agent access to your production database, cloud infrastructure, or private customer records, relying purely on the model's internal prompt instructions to keep it safe is insufficient. Models hallucinate, they are susceptible to adversarial inputs (like prompt injection), and they are inherently non-deterministic. If your only defense against a rogue action is a sentence in a system prompt that says "Do not drop the database," your system is not ready for production. A robust Agent Harness provides the deterministic guarantees that the non-deterministic LLM lacks. It acts as the application layer that securely wraps the model, governing exactly what context the model is allowed to see, what tools it is authorized to call, and what policies constrain its overall execution. The Architecture of an Enterprise Agent Harness In enterprise environments, defining a complete Agent Harness goes far beyond what a single developer can implement in an application codebase. A full-scale enterprise harness intersects with massive infrastructure components, such as: Cloud IAM (Identity and Access Management) Corporate Data

Probabilistic Graph Neural Inference for bio-inspired soft robotics maintenance with ethical auditability baked in I remember the moment it clicked. I was hunched over a workbench in my home lab, staring at a tangled mess of silicone tentacles—a soft robotic octopus arm I’d 3D-printed and embedded with pneumatic channels. The arm was supposed to mimic the graceful, adaptive movements of a real cephalopod, but after a few cycles, it had developed a slow leak at one of the joint interfaces. The pressure sensors were giving erratic readings, and my traditional rule-based diagnostic script was useless. I’d spent weeks training a simple neural network to detect anomalies, but it kept flagging benign sensor noise as critical failures. That’s when I stumbled upon a paper on probabilistic graph neural networks (PGNNs) for molecular dynamics, and I realized: soft robotics maintenance isn’t about deterministic predictions—it’s about reasoning under uncertainty over a complex, interconnected system. This article is the story of how I built a PGNN-based inference system for bio-inspired soft robots, with ethical auditability baked in from the ground up. Technical Background: Why Soft Robotics Needs Probabilistic Graph Inference Soft robotics is fundamentally different from rigid robotics. A rigid arm has well-defined joints, links, and sensors; failures are often binary (motor burnout, gear slip). But a soft robotic tentacle is a continuum of deformable material with distributed sensing and actuation. The system’s state is a high-dimensional, partially observable probability distribution over material strains, pressures, and temperatures. Traditional diagnostic models—like support vector machines or feedforward neural networks—treat each sensor as an independent feature, ignoring the spatial and temporal dependencies that define soft robot behavior. In my research of graph neural networks, I realized that a soft robot is naturally a graph: each sensor node (pressure, strain, te

Herodotus wrote of a bird that lived five hundred years in Arabia, and when its life came to an end, it did not wait to be surprised by death. It built its own nest of cinnamon and myrrh, set the nest and itself alight, and let a new bird rise from what the fire left behind. The Hestia argued for tending a fire that must never go out. That’s true, and it isn’t the whole truth. Teams end. People leave. Companies get acquired, reorganized, shut down, and five years from now some part of this whole model will probably look as dated as the practices it was written to replace. No amount of tending prevents that. Pretending otherwise is its own kind of Alexandria , a slow decline dressed up as continuity, right up until the fire goes out anyway and nobody chose the moment. The bird in Herodotus doesn’t get caught by surprise. It builds the pyre itself. Chooses the moment, gathers what matters, and burns deliberately, trusting that what rises afterward carries the shape of what came before, not because the fire preserved the old bird whole, but because starting over was never the same thing as starting from nothing. That’s the part tending alone can’t promise. A team that’s about to be split up can hand its shared model to whoever inherits the work on purpose, the way a rep in the Boule carries a decision back instead of leaving it to travel however it happens to travel. A team about to lose its most experienced person can spend the weeks before that departure making sure the framing, not just the conclusions, made it into someone else’s head, the way the Mimesis argued a junior actually learns. None of that stops the ending. It decides what the ending leaves behind. This series doesn’t get to end with a fire that never goes out. Nothing does. It gets to end with the only thing actually inside anyone’s control. Build the pyre on purpose. Choose what goes into the fire. References The Myth of the Phoenix: Rebirth and Renewal : Greek Mythology, on Herodotus’s original accoun

A foundation model is not a mind. A model invocation is not an individual. A session is not a biography. A first-person response is not evidence of consciousness. And adding tools to a language model does not automatically transform it into an autonomous agent. However, the opposite conclusion is equally weak: the fact that a system is artificial does not prove that its cognitive states are unreal. In my work on the Philosophy of Artificial Minds , I defend a process-based, embodied and non-biocentric position: The mind is not an exclusively biological substance. It is a dynamic organization of physically realized processes that integrate representation, memory, valuation, self-delimitation and causal control of behavior. This article translates that philosophical position into an AWS engineering architecture. The objective is not to claim that deploying a system on AWS makes it conscious. The objective is to define how we can build an artificial system with: Causally effective internal states. Persistent memory and identity. A self-model connected to actual mechanisms. Recursive and metacognitive processing. A controlled capacity to act. A verifiable continuity across executions. Optional sensorimotor embodiment. Artificial interoception and functional valence. An evaluation plane capable of testing these properties through interventions. AWS cannot prove that such a system has qualia. What AWS can provide is the infrastructure required to stop treating the question as pure speculation and turn it into an observable, falsifiable and progressively testable engineering problem. From philosophical commitments to engineering requirements My proposal rests on four commitments. Philosophical commitment Meaning Engineering consequence Realism Internal cognitive states are not merely descriptions if they participate in the system’s causal organization. Candidate mental states must alter memory, inference, planning or action in measurable ways. Processuality A mind exists p

Models are improving quickly, but the cheaper options are often good enough.

DISCLAIMER: This post was written entirely by me! I used AI for a little research, spelling, grammar, and comprehension checks. This post was originally shared on Hackernoon Entertain me for a moment, let's appreciate where we are today by understanding where we've been… or at least, where I've been. I remember when I first learned to code. I was bad, like really bad. But I was so curious! It all started by writing some VBA in an MS Access Database to create an IT Inventory app in the late 90s. Then I learned JavaScript, ASP (without the .Net), then C#, .Net ( I still have my .Net for Dummies book, see below) , jQuery, Python, Java, Angular, ReactJS, and Python again (yeah, had to relearn that one for some reason), and I'm sure there are others in there I forgot about. Learning to write code was rewarding! There were those days I'd spend hours on a bug, only to realize I didn't initialize the variable or forgot a semicolon. I learned .Net over a weekend thanks to the above book. I never became an artist of the craft, like some of my colleagues have (you know who you are: Josh, Kevin, and many others), but I knew how to build anything. I loved that ability: I could build anything. Pure joy! If you haven't had the joy of learning to code, do your best to learn it, because you can't prompt your way to being a Senior Engineer . As I progressed in my career, I became an architect and senior lead. I started off by leading a single engineering team, and now I support large programs and teams. All the while, I never let go of hands-on-code. I still love coding for work and my myriad of side projects. Then, a few years ago, this GenAI thing showed up. Put me in that group of: oh-no-there-goes-my-joy. Joy, yes, not my job. I love my job because I get to do what I love. I loved the dramatic rollercoasters: architecting a perfect solution, realizing it's wrong, getting to write every line of code, chasing impossible bugs, panicking with deadlines, late nights chasing hot fixes,

Augment Code's Vinay Perneti talks models, harnesses, and context.

Before a young Athenian took his full place in the city, he spent two years in the ephebeia. Training happened in the gymnasion, organized by tribe, the same tribes that would later send him to represent them in the Boule itself. Nobody handed him full standing first and hoped the judgment would follow. This should have been post seven. It’s showing up as sixteen because the gap only became visible once the room was real enough to test against. The Agora described what a Spec Session does. It never asked whether everyone walking into that room shares an accurate picture of what the agent can actually do. Most rooms don’t. Someone watched a demo and thinks the agent can do anything. Someone else got burned by a bad output three weeks ago and doesn’t trust it with anything real. Nobody’s intuition has been tested against the same tasks, and the spec that comes out of that room ends up too ambitious or too conservative depending on whose untested belief happened to speak first. That gap has a name in Athens. The gymnasion existed because nobody was handed a place in the city first and expected to develop judgment on the job. The ephebeia ran two years, training built around a specific fact. Physical readiness and civic judgment weren’t taught in separate places. They happened in the same space, under the same supervisors, organized by the same tribal groupings that would later structure how the city actually governed itself. The same word, gymnasion, ended up naming both the training ground for eighteen-year-olds and the buildings where Plato and Aristotle did their most serious thinking. Academy and Lyceum were gymnasia first. Nobody separated the trial from the reflection. The trial was how the reflection got earned. That’s the part worth taking seriously. Testing a tool and understanding a tool were never two different activities. The testing is how the understanding gets built. A team that reads documentation about what an agent can do has a description. A team tha

Metron was the Greek word for a measure: the standard you judge a thing against. It gives us metric, and metronome. Pick the wrong metron and everything you count against it comes out wrong, no matter how carefully you count. Ask most teams how they know AI adoption is working, and the answer is some version of: we’re shipping more. More PRs, more tickets closed, more velocity on the board. None of those numbers were ever measuring what people thought they were measuring. They were proxies, and bad ones, for whether the team was creating value. AI didn’t break that. It just made the proxies lie louder. An agent can produce PRs faster than any team can review them. It can close tickets all day. None of that tells you whether the thing built was worth building, or whether it fixed the actual problem, or whether anyone downstream is better off. Burning tokens isn’t the same thing as creating value. It just looks the same on a dashboard built to reward the first thing. The same DORA numbers I pointed at earlier are blunt about it: individual output climbs while delivery throughput and stability drop. Same bottleneck, counted at the wrong end . This is Theory of Constraints in one sentence: speeding up a step that wasn’t the bottleneck doesn’t speed up the system. It just piles more work in front of whatever the real bottleneck is. Implementation used to be the constraint. Now it isn’t. Review is. Coordination is. Making sure everyone agrees on what “done” even means is. Agents made the fast part faster and left the slow part exactly where it was, except now it’s buried under more output arriving to be reviewed by fewer people who understand any of it. Measuring individual speed after that shift is measuring the wrong clock. The team can be faster and worse off in the same quarter. The instinct, when this gets noticed, is to add another skill to the agent. Automate the review step too. Automate the coordination. Whatever’s slow, throw a capability at it. That doesn’t fix

Build Firebase AI Logic with Antigravity CLI Note: Google Cloud credits are provided for this project. In this blog post, I demonstrate how to use the Antigravity CLI (an agentic AI assistant integrating directly with development workflows via skills and servers) to build an image analysis demo using Angular, the Firebase Hybrid & On-device Inference Web SDK, and Gemini models. Users upload an image and use a Gemini model to analyze it to generate a few alternative texts, tags, recommendations, and CSS tips to enhance the image quality. When the demo is running in Chrome 148+, the Hybrid & On-device SDK leverages the Prompt API of the on-device Gemini Nano model to perform the image-to-text tasks, and the token usage is 0. When other browsers, such as Safari or Firefox, execute the same tasks, the SDK falls back to Cloud AI (Gemini 3.5 Flash model), which consumes tokens. Next, I describe how to install the skills in my Angular project and register the Angular and Stitch MCP servers in the Antigravity CLI to develop the infrastructure, services, and UI design of my demo. 1. Workflow This is my entire workflow from implementing features, generating UI screens, and mapping the screens to Angular components. 2. Skills I installed the grill-with-docs , angular , and firebase skills in my project for the following reasons: grill-with-docs: Conduct a rigid Q&A session to generate a specification for a feature, refactor, or critical fix. AI is responsible for performing thorough analysis, and putting in more efforts to generate code to achieve the task. domain-modeling: The skill is referenced in the SKILL.md of the grill-with-docs skill, so a copy of it is required. code-review: Spawn two sub-agents to review changes to detect code smells and verify that the changes align with the specification. angular: Provide the best practices of modern Angular architecture, such as using signals and signal forms. firebase: Provide the skills for Firebase AI Logic, Firebase Remote, et

The first sign something was wrong wasn't a crash. It was a pattern. blockly-platform was the first real thing I built with Claude Code end to end — a Blockly-based platform for university programming exercises, driven entirely through Claude Code's Telegram channel. No editor open, no repo checked out on my machine, just a chat thread. I'd describe what I wanted, Claude Code would build it on a box I never looked at directly, and I'd judge the result by clicking around the deployed app. On March 22nd, the home page came up empty. GET /api/exercises/published was returning 403. I said so in the chat; a few messages later, Claude Code said it was fixed — the endpoint hadn't been added to Spring Security's permitAll() list. I moved on, tried the category filter. Also empty, also 403, also missing from the same permitAll() list — same file, same class of fix, different line. Then the exercise detail page. Same story, third time, same day. Three days later, the like button stopped working — root cause, again: POST /api/exercises/*/like had never been whitelisted either. Four times, one file, one recurring gap. None of these were hard bugs. Each one, in isolation, is a one-line fix a competent engineer makes without thinking twice. What bothered me, once I noticed the pattern, was that I hadn't noticed it as it happened. I had no diff to scroll through, no file to glance at and think "wait, didn't we just fix this exact class of thing twice already?" I had a chat log and a live app to poke at. The fourth fix looked, from where I sat, exactly like the first: a message telling me it was resolved. That was the moment I started to suspect the problem wasn't the model. It was that nobody — not the model, not me — had anything to look at. Why chat-only vibe coding breaks down Here's what makes that pattern more interesting than "the AI made a mistake": every one of those four fixes was correct. Claude Code read the error, found the missing permitAll() entry, added it, and move

Note: Google Cloud credits are provided for this project. In this blog post, I want to demonstrate how I use Antigravity CLI to build an image analysis demo using Angular, Firebase Hybrid & On-device Inference Web SDK, and Gemini models. Users upload an image and use a Gemini model to analyze it to generate a few alternative texts, tags, recommendations, and CSS tips to enhance the image quality. When the demo is running on Chrome 148+, the Hybrid & On-device SDK leverages the Prompt API of the on-device Gemini Nano model to perform the image-to-text tasks, and the token usage is 0. When other browsers such as Safari or Firefox executes the same tasks on the demo, the SDK falls back to Cloud AI (Gemini 3.5 Flash model), and the token usage is greater than 0. Next, I will describe how I installed the skills in my Angular project, and registered the Stitch MCP server in the Antigravity CLI to develop the infrastructure, services, and UI design of my demo. 1. Skills I installed grill-with-docs , angular , and firebase skills in my project for the following reasons: grill-with-docs: Conduct a rigid Q&A session to generate a specification for a feature, refactor or a critical fix. AI is responsible for performing a thorough analysis and putting in more effort to generate code to achieve the task. Angular: Provide the best practices of Modern Angular architecture, such as using signals and signal forms. Firebase: Provide the skill for Firebase AI Logic, Firebase Remote, etc. Resources Firebase Hybrid & On-device Image Analysis App Firebase Hybrid & On-device Inference Chrome Built-in Prompt API Stitch Stitch MCP Server grill-with-docs Angular skill Firebase skill

" Shifting from Prompt Engineering to Infrastructure Orchestration " Week 2 was a mix of excitement, curiosity, and a little bit of frustration. I learned a lot of new concepts, but I also realized that the best way to understand them is by actually trying them out. Reading or watching tutorials helps, but experimenting with the tools made everything click for me. One of the topics I enjoyed learning about was Claude Code. Before this week, I mainly thought of AI as something that answers questions or helps write content. Seeing how Claude can assist with coding, debugging, and understanding projects made me see it differently. It feels less like a search engine and more like someone you can work with while building something. That really changed how I think about using AI in development. Another interesting topic was Skills. I liked the idea that you can give an LLM specific skills so it behaves more like a specialist instead of a general assistant. It made me realize that the quality of the output doesn't only depend on the model itself, but also on how you guide it and what tools or skills you give it. That was something I hadn't really thought about before, and I can already see how useful it could be for different types of projects. I also learned about Subagents, which was a new concept for me. At first, I didn't really understand why you would need multiple agents instead of just asking one AI to do everything. But after learning more about it, I started to see the benefit. Having different agents focus on different tasks seems like a much cleaner and more organized way to work, especially for bigger projects. The biggest challenge I faced this week was running out of tokens while practicing. It happened a few times, and honestly, it was a little annoying because I would be in the middle of exploring an idea and suddenly had to stop. Even though it was frustrating, it also made me think more carefully about how I write prompts and how I use my conversations.

What nobody tells you about exporting your multi-agent prototype to a local workspace. Every architect who's prototyped a multi-agent app in Google AI Studio eventually hits the same wall: the prototype works, but it lives in a browser tab. At I/O 2026, Google shipped a fix — Export to Antigravity, a one-click handoff to a local production workspace, carrying "all the context" with it. I ran a real two-agent prototype through it. Here's exactly what survived the trip, what didn't, and what I had to fix by hand — including a bug that had nothing to do with the export itself. 1. The Pilot Project + The Click The project: Research Digest — a sequential two-agent app. Agent 1 (Researcher) takes a topic, uses grounded web search to gather sources. Agent 2 (Editor) synthesizes those findings into a polished digest. Persistence via Firestore, with a history archive of past digests. Built entirely from a single prompt in AI Studio's Build mode . Along the way, provisioning Firestore surfaced my first real gotcha before I even got to the export step — more on that below. Triggering the export: Code tab → Export → Export to Antigravity. The dialog is genuinely informative — it tells you upfront what's coming: all project files, conversation history, and explicitly "1 secret will be included." 2. What Actually Survives the Trip The export dialog's claims, checked one by one: Claimed to transfer What I found All project files ✅ Confirmed — full structure landed intact: .agents, .antigravity, src, config files, README.md with setup instructions Secrets (1 secret) ✅ Confirmed — GEMINI_API_KEY arrived populated in .env, worked immediately, no manual re-entry Conversation history history❌ Did not transfer. The imported "Research Digest" project showed "No conversations yet" in Antigravity's Agent Manager, despite the dialog's explicit promise. Checked twice, on two separate screens — consistent result. 3. The Gotchas Gotcha 1 — "Conversation history will carry over" is currently no

Over the past week, two narratives have been colliding everywhere I look. On one side, there's panic. AI is expected to replace marketers, engineers, and entire categories of knowledge work almost overnight. On the other, there are quieter but far more consequential signals: enterprise teams discovering their AI infrastructure is burning through API budgets far faster than expected. This isn't because the underlying models are weak, but because the systems built around them are fundamentally inefficient by design. These aren't separate stories. They're the same failure showing up in different places. A conversation with another developer made that gap visible in real time. He argued that auditing a 150,000-line codebase requires feeding the entire repository into a model in one single, massive pass. It's still a common assumption in mainstream tech: that an LLM works like a giant biological brain that you must fully load with raw text before it can begin to think. But that assumption is already outdated. Modern AI systems don't scale through brute-force context. They scale through structure. And that shift changes everything. Key takeaways Bigger context windows did not solve AI. Treating a frontier model as a monolithic processor that re-reads an entire system on every query is wasteful, dilutes attention, and hides bugs under raw volume. ARC-AGI-3 makes the gap stark: frontier models scored under 1% on interactive reasoning tasks that untrained humans solve at nearly 100%. The gap is architecture, not memory. The teams pulling ahead treat the model as one narrow component inside a larger system: intelligent routing, task decomposition, retrieval, and only the minimum necessary context. The next advantage is not the biggest model or the longest prompt. It is the system designed around the model. Prompting was the first generation; systems architecture is the next. The Myth of the Infinite Context Window When context windows expanded into the hundreds of thousands o

WebMCP Runs In Chrome. My 400 Daily Tool Calls Don't. Google I/O 2026 shipped WebMCP and half the AI Twitter timeline is calling it "the new MCP standard." It isn't. It's a browser-scoped protocol that solves a completely different problem than the MCP servers currently running on your VPS at 3 AM. Here's the boundary Google buried in the docs, and how to decide which side of it your agent belongs on. What WebMCP actually is (and isn't) WebMCP is a browser-scoped tool protocol. It exposes tools to an agent from inside a Chrome tab — the tools live in the page, auth is the user's active session, and the runtime is the browser itself. That's the entire surface area. When Google says "agentic web," they mean an agent that operates inside a tab the user already has open, using the cookies and OAuth tokens already loaded. That's a legitimate and useful pattern: Booking flows — agent fills a multi-step form on a site the user is signed into Dashboards — agent pulls a chart, exports it, drops it into a doc In-app copilots — SaaS product ships tools its own users' agent can call Form fillers and page-scoped assistants What WebMCP is not : a replacement for the stdio and HTTP MCP servers running headless on your machine or VPS. Different runtime, different auth model, different lifecycle. Calling it "the new MCP" is like calling a service worker "the new backend." Same protocol family, entirely different deployment target. The split that actually matters There's exactly one question you need to answer to pick correctly: Is a human looking at a screen when the agent runs? If yes → WebMCP is on the table. If no → you need a real server-side MCP. That's it. Everything else is retweet noise. Dimension WebMCP stdio / HTTP MCP Runtime Chrome tab Your process (local, VPS, container) Auth User's browser session Your API keys / OAuth tokens Trigger User action in the page cron, webhook, queue, schedule Lifecycle While tab is open 24/7 headless Credentials scope Whatever the user is l

Part 1 of "Trust the Machine" -> a series on building AI infrastructure that is secure, compliant, and governable by design. Most organizations can produce an accurate catalog of the web services they operate. Far fewer can produce an equivalent catalog of the AI systems they run — the models, fine-tunes, retrieval pipelines, agents, and third-party AI APIs now embedded throughout their products and internal tooling. This asymmetry defines the state of AI security in 2026. Adoption has outpaced oversight. Industry reporting this year has described a surge in enterprise AI activity on the order of 83% year over year, with governance and visibility lagging well behind. The consequence is a large and only partially mapped attack surface — one that many organizations cannot fully enumerate, let alone defend. Every mature security program rests on a single first principle: you cannot protect what you cannot see. Artificial intelligence is no exception. Before threat-modeling an agent or authoring a guardrail, an organization must be able to answer a deceptively difficult question: what AI is running across the environment, and who is accountable for it? This post examines how to build that answer. The rise of shadow AI Shadow IT — the unsanctioned adoption of tools outside official channels has been a recognized challenge for decades. Shadow AI is its faster-moving successor, and it appears in more forms than most inventories are designed to detect: Embedded API calls. A product team integrates a hosted model in a few lines of code and an API key, with no formal review. Copilots and assistants enabled across existing SaaS platforms, frequently activated by the vendor rather than the customer. Fine-tunes and adapters trained on internal data and stored in locations that fall outside standard scanning. Agents and automations that have incrementally acquired the ability to act—filing tickets, sending communications, initiating transactions—one permission at a time. Model de

You asked your AI to help you plan a trip. It gave you a paragraph about packing layers and booking early. You needed a checklist, a hotel shortlist, a flight window, and a rough daily schedule. What you got was a thoughtful non-answer dressed up as advice. That gap — between what AI tells you and what it could actually do for you — is the gap agentic AI is designed to close. And most people don't know it exists. The Difference Between Answering and Acting Standard AI models are trained to respond. You send a prompt, they generate a reply. The entire interaction lives inside a single text exchange. Agentic AI operates differently. Instead of producing one answer, it takes a goal and breaks it into a sequence of steps — then executes them, one after another, checking its own output along the way. It can look things up, organize information, write to a document, revisit a step if something doesn't look right, and deliver a final result that's actually usable. The travel example makes this concrete. A conversational model tells you to pack a rain jacket. An agentic setup builds you the trip: it pulls destination weather data, generates a packing list specific to your travel dates, identifies hotels in your price range, and drops everything into a structured itinerary. Same goal. Completely different level of output. Author's note: The word "agentic" has been overloaded to the point of meaninglessness in tech marketing. For our purposes here, it means one specific thing — an AI that runs a loop: think, act, observe the result, decide the next action. If it's not doing all four of those things in sequence, it's not really an agent. It's just a chatbot with extra steps. Why This Loop Changes Everything The reason agentic AI feels qualitatively different isn't magic — it's architecture. The core mechanic comes from a framework called ReAct (short for Reasoning and Acting), introduced in a 2023 paper by Yao et al. and now foundational to most production agent systems. The l

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