og.rsts.dev

Writing

159 posts · 10 series · since 2026

2026

159 7 min You Can't Forward the Understanding Pasting a model's answer into a pull request, a design doc, or a Slack thread feels like leverage. It is a quiet trade: you keep the credit and hand away the competence. The work the tool cannot do for you, reading and verifying and re-expressing, was the work that kept your judgment sharp. 158 12 min Trunk-Based Development Finally Came for the Database A builder's playbook for trunk-based development on the database: branching versus migration-as-code, expand/contract in real SQL, online changes that don't lock, and the CI gate, across Azure SQL, Postgres/Neon, Snowflake, and Databricks. 157 11 min The Chat Box Is the Cheap Part A build guide for self-hosting an enterprise AI chat front-end when a Teams bot can't promise US data residency. Entra at the perimeter with zero app auth code, Azure OpenAI pinned to a US region, and the config to ship it. 156 9 min The Harness Is Becoming the Runtime A builder's read on durable execution moving into the agent harness: wrap your agent in Temporal versus owning a durable harness, what LangGraph's checkpointer actually guarantees, Pi Durable's replay policy, and how to choose. 155 12 min Opacity Was the Moat Cryptographers settled it in 1883: a system that depends on secrecy was never secure. The rest of software coasted on opacity anyway, because understanding a system you did not build was expensive. That cost has collapsed, from AI decompilers to spec recovery for legacy code, and with it the last pretense that a black box is a wall. Here is what that breaks, what it exposes, and what survives. 154 9 min Salesforce Is Writing Your Exit Clause A platform confident in its future does not decompose itself into endpoints, give away the data layer, and hand you React. Salesforce is unbundling itself because the human-seat era its footprint was built for is ending. Agents don't log in. The real question is which two layers survive the transition, and whether you own them. 153 9 min Leaving nCino Won't Set You Free nCino runs on Salesforce, so the exit does not lead outside. For most banks it leads one floor down, into deeper dependence on the platform you were already renting. There are three doors out of nCino. Only one of them actually sets you free, and it is the one almost nobody takes. 152 12 min The Exit Clause Your Architecture Is Missing Vendor lock-in is a weaker trap in the AI era than the renewal quote assumes. The leverage was never the platform, it was the cost of understanding your own estate, and that cost just collapsed. A practical path off MuleSoft, and why every renewal needs an exit already written. 151 8 min The Economics of Renaming AI Musk renamed SpaceXAI to SpaceXSI. Benioff flipped AIForce to SIForce. A September 29 executive order told federal agencies to stop writing 'artificial intelligence' and start writing 'super intelligence.' As an engineer I know what that is: a find-and-replace. Here is what the rename actually ships, what it doesn't, and why the gap between those two is where the money gets made or lost. 150 8 min Don't Build the Hippocampus Last week I argued the enterprise second brain needs a body, and that the recall organ was the hippocampus. The recall organ is already built: Hindsight, a mature, widely-adopted, company-backed open-source agent-memory system. So the question stops being how to build the hippocampus and becomes why you would. This is the build-versus-buy case: reuse the maintained commodity organ, and spend your scarce build budget on the parts only the lakehouse can grow, the governed business data, certified meaning, and audit plane that are irreducibly yours. 149 13 min A Second Brain Needs a Body Your second brain is brilliant and it cannot come to work. Not because it doesn't know enough, but because it has no body: no permissions, no immune system, no way to serve anyone but you. This is the personal second brain scaled from the individual to the enterprise, and the argument that the lakehouse is the body the cortex has been missing. Markdown is the authored meaning; Databricks is the governed substrate that keeps it alive at scale. 148 13 min From Medallion to Meaning: Databricks as an Enterprise AI Context Layer A data lake doesn't ground anything. Hand an AI agent a pristine Gold table and it still has to guess what the number means. This is how Databricks turns storage into meaning: the lakehouse and medallion architecture as the foundation, Unity Catalog and metric views as the governed semantic layer, and Vector Search plus Mosaic AI as the grounding path for agents. The stack, what each layer actually does, and where the lock-in hides. 147 9 min One Gateway, Three Kinds of Traffic: What the POC Has to Prove A single gateway that governs classic APIs, LLM calls, and agent-to-agent traffic from one control plane is an easy demo and a hard proof. The demo shows the happy path once. The POC has to show the control plane actually enforces under load: native host identity on every lane, private network paths, token counts that reconcile, policies that halt rather than log, and end-to-end correlation IDs across all three kinds of traffic. Here is what a serious proof-of-concept has to establish, and how to test each claim. 146 6 min IT Is the Blueprint Every post in this series built one pillar of a larger machine. The AI-first enterprise is an operating system with four of them: Context, Prompt, Model, and Tools, each solved from the individual up to the enterprise. The reason IT gets to go first is that IT already runs the blueprint. The governed lifecycle isn't an aspiration in software, it's Tuesday. This is the case that IT is the reference implementation the rest of the business copies, and the honest account of what it still can't do for you. 145 8 min Let the AI SRE Read Everything, Write Almost Nothing The AI SRE is sold as the agent that fixes production before you wake up. It is also another non-deterministic actor with a hand on prod, the exact thing you were warned about. The category bundles two jobs under one name: investigation, which is a read and a safe yes, and remediation, which is a write and needs a gate. Split the verbs and the value is real. Bundle them and you have automated a new way to break prod. 144 6 min One Word Moves the Control The governed lifecycle has one weak joint, and it is the review gate. Code review works partly because the machine runs the test and tells you, objectively, that something broke. For prose where meaning is the payload, one changed word can move a control, and no test catches it. This is where everything-as-code stops being a mechanical win and starts demanding human judgment, and why automating a broken review just scales the mistake. 143 7 min Koa Isn't the Best Model. It's the Best-Placed One. Salesforce's first CRM reasoning model trails frontier models on the independent benchmark and near-tops Salesforce's own. That contradiction is the story. Koa is not a capability play, it is a placement play: reasoning post-trained for CRM actions, running inside the vendor's trust boundary, next to the data and identity it needs. Judge a vertical model on placement and action accuracy, not on a leaderboard it was never built to win. 142 6 min Grounding Is a Memory Problem Everything-as-code builds the semantic layer of context, the governed meaning of the business. It does not build the other half. Grounding an AI turns out to be a memory problem, and the answer has always been the second brain, at every scale from a person's notes to an enterprise knowledge base. Here is the practical shape of memory at the individual, team, line-of-business, and enterprise level, and why every framework keeps rediscovering what neuroscience still can't fully explain. 141 6 min Version Control for Data Git is the source of truth until the artifact is a terabyte of unstructured content or a billion-row table, and then it isn't. The principle survives the medium: text lives in git, unstructured content lives in a lineage-tracked lake, structured data lives in a lakehouse whose table-format time travel is literally version control for data. Same invariant, three substrates, chosen by the shape of the thing. 140 8 min Compile the Policy A risk standard today is a document in a shared folder with the version typed into the footer, and the enterprise calls that governance. Walk it through the software chain instead: authored from source, rendered on build, tested as policy-as-code in a throwaway sandbox, versioned immutably in a registry, promoted through sign-off gates into the system it governs. None of it is new. It is the pipeline software has run for years, pointed at the document. 139 10 min The Meter Is the Feature Microsoft made Copilot Cowork generally available on consumption billing, then in September shipped spending policies, alerts, and approval workflows in the admin center. Read as a feature wave it looks like productivity. Read as architecture it is the tell that agentic work is now its own runtime tier, and the real decision is where you place that tier and how tightly it couples to your context layer. A practical guide to running the agentic-work layer through Copilot Cowork, Claude direct, or Salesforce Agentforce, decided on placement and exit cost rather than who owns the meter. 138 7 min The Architecture Belongs in the Repo Most enterprises review architecture once a year in a slide deck nobody can diff. Put the model in the repo instead, as ADRs and C4 and contracts, and architecture gets what code has: a change ticket, a review that gates the merge, a test that fails when a decision violates a standard. Governance stops being a meeting and starts being a pull request. 137 9 min Jev Decides. You Still Own the Call. The most interesting model of the month never writes a sentence. Jev hands back a typed verdict and a confidence number in milliseconds, for a fraction of a cent. The new primitive is a calibrated, structured, auditable decision, not another chatbot. Here is where it actually fits in a regulated stack, how it drops into a Java shop, and the enterprise gate the cheap-and-fast pitch still has to clear. 136 10 min You Bring React. Salesforce Keeps the Runtime. Salesforce Multi-Framework lets you build in React and embed it inside the Lightning shell, and it is the inverse of the year's decompose-outward story. The open layer is framework choice. The durable layer is the runtime, the data access, and the governance, and those stayed Salesforce's. Here is the honest case for the developer-experience win and for the control a regulated shop is quietly trading away. 135 8 min The Domino That Already Fell Infrastructure was the first enterprise deliverable to leave the click-console for a governed lifecycle, and it worked. Declared desired state in git, a reconciler that drives reality toward it, drift detection, rollback by revert. This is the domino that proves the pattern. The rest of the series is asking why we stopped at infrastructure. 134 16 min AI Replaces the UI. What an Architect Actually Takes Home Dreamforce 2026's headline was AIforce: AI replacing the interface itself. Underneath it sits an old argument with two right answers. The customer and the banker want the friction gone. The architect has to keep control of identity, data, and cost. Here is the honest case for each side, grounded in what the keynotes actually said, and how the control plane lets you say yes to both. 133 12 min The Model Decides, the Code Controls The post-Dreamforce answers to the three questions every regulated shop is asking about agents: where the model decides and where the code does, why shipping is the real work and not the demo, and how bounded economics give a CFO a number to underwrite. The governance is what gets agents into production, not what holds them back. 132 10 min Data 360 Is a Reference System, Not a Database What the Data 360 developer bootcamp actually taught me: its real power is as a semantic layer over data you never move, a reference system rather than another place to manage records. Two places that reframe pays off: marketing segmentation, and unifying one person's fragmented logins across every portal a bank runs. 131 12 min I Already Solved This Every enterprise governance problem I look at lately triggers the same déjà vu. Across 35 years writing software I watched one class of deliverable earn a governed lifecycle the hard way, through Agile, TDD, DevOps, GitOps: source of truth, change ticket, review before merge, gated pipeline. Every other enterprise artifact is still sitting where code was before those inflections. AI is what finally makes running that loop cheap enough to point at all of them, and the center of the AI-first enterprise is a source of truth you can version. 130 12 min Draw the Boundary Before You Buy Deeper I'm flying out early for the Data 360 developer bootcamp, but the bootcamp is the cheap part. The hard question at Dreamforce is where the boundaries are: on platform versus off, data versus integration versus action, and whether source-driven development is enforced or just available. Here are the questions I refuse to leave without answering. 129 8 min Who Judges the Judge? The industry's answer to authorizing an agent action is to put another model in front of it as a judge. That works right up until the judge and the actor share a blind spot, which, being the same model family, they usually do. Correlated judgment is the failure the pattern is built not to see. 128 8 min Least Privilege Was for Humans. Agents Need Least Agency. Access control was built for a bounded actor who does one thing per credential. An agent with a valid token can do anything the token allows, in any combination, at machine speed. The control you actually need is not least privilege on the credential, it is least agency on the behavior. 127 9 min SR 11-7 Was Written for Models That Don't Argue Back The supervisory guidance every bank uses to govern models assumes a model is a fixed function you validate once and monitor for drift. A generative agent is non-deterministic, self-modifying in context, and calls tools. The framework you are required to use has no box for it, and the examiner still expects one. 126 10 min Stop Tuning the Agent. Build the Environment. The agent harnesses converged into a commodity. Whatever vendor you pick, the loop is the same. The hard part was always the thing nobody demos: the environment the agent lives in, builds in, and ships from. 125 10 min Don't Vibe Your Infrastructure Vibe coding turns an idea into a diff. VibeOps is everything after it, minus the human who used to read every line. Build infrastructure as a reviewed IaC artifact through an agentic development lifecycle. The chat is the new console for operating what is already built, not a shortcut around delivery. 124 8 min The Craftsman's Goodbye I spent more than thirty-five years getting good at something a machine now does in seconds. The grief is real, and skipping past it is dishonest. But the years weren't spent on the craft. They were spent building the eye. 123 10 min Identity Isn't Intent In an agent-driven enterprise the caller and the model are both replaceable. What you defend is the governance that answers three questions in order: who runs it, did anyone ask, and is the effect allowed. Identity solves the first. The gap is the second. 122 9 min Be Vague on Purpose The instinct that made you a great requirements author is making you a worse prompt author. Over-specify the how and you cap the model at your own imagination. In a bank, the trick is knowing exactly what you must never leave vague. 121 11 min The Headcount Math Is Wrong Enterprises buy agents to cut headcount. The subtraction never works, because agents don't remove work, they move it. The bank that already ran this experiment with ATMs should recognize the pattern. 120 9 min You Can't Edit What You Can't Write Reviewing, auditing, and defending code to an examiner all presuppose the ability to build it. When institutional competence erodes, that's not personal regret. It's concentration risk on a three-year timeline. 119 10 min The Machine Has Better Discipline Than Your Team In a regulated bank, discipline is compliance. The agent that writes the test, fills the PR, and documents the decision every single time is a better audit citizen than your best engineer at 5pm on a Friday. 118 12 min Salesforce Put Its CRM Inside Claude Salesforce and Anthropic shipped Salesforce in Claude, a plugin that lets sellers act on live CRM data without opening the app. Strip the demo away and it is a bet about which layer holds the value: the interface Salesforce built, or the governed data underneath it. Here is what it is, what each side gets, and the case for staying skeptical. 117 10 min Microsoft's Claude Looks Cheaper Until You Read the Prerequisite Microsoft says its Copilot Cowork runs Claude 30 to 40 percent cheaper than going direct. The number is real. It also assumes you already bought a thirty-dollar-a-seat license you may not need. Here is the actual cost math for putting Claude to work over Microsoft 365, and what you trade away going direct. 116 10 min Claude Buddy Stopped Being an Extension and Became an Ecosystem In April, Claude Buddy became an extension of a personal-AI framework. Today it became four plugins that, together, vendor almost none of their own code. The PAI-to-LifeOS migration shows why that was the right bet. 115 10 min Core-Agnostic Still Needs a Core Fiserv's agentOS puts agents inside its stack. FintechOS puts them above any core. They are opposite bets on who owns the layer above the ledger, and both land on the same dependency: a modern real-time core. On the Fiserv stack, that means Finxact, and almost nothing else. 114 11 min Banks Don't Need Another Agent. They Need a Context Layer. Only 16% of enterprise AI deployments are real agents, and Gartner expects 40% of agentic projects dead by 2027. The layer that decides which ones survive isn't the model. It's the governed context underneath, and inside a bank that same layer is the audit trail. 113 6 min The Best Decision My Daughter Made Was Betting on Herself My daughter is a computer science student who could have taken any safe path. Instead she started her own company and has been building it from scratch, in public, ever since. A proud father on what it looks like when someone bets on herself and does the work. 112 10 min The Data Was Always the Problem The verdict from the Fiserv Forum floor: the agent story is real, but announced is not turnkey, and underneath the keynote sits the problem banks never solved. Fiserv and Salesforce are fighting for the same agent surface on opposite stacks, and both are really fighting over who becomes your data's context layer. The last dispatch of three. 111 8 min agentOS Has to Eat Fiserv's Own APIs Fiserv's agentOS promises agents that do real banking work. That promise inherits the reliability of every API and core beneath it, and the hard test isn't the demo at 2pm, it's the unattended agent at 11pm while the core is closing the books. An architect on dogfooding, the end-of-day problem nobody demos, and whether the plumbing ships on time. 110 8 min They're Already Copying Your Encrypted Data The post-quantum threat is not a future decryption event. It is a present-tense collection one, and the traffic being copied off the wire tonight is the traffic an adversary plans to read a decade from now. For any institution with a long data-sensitivity horizon, the migration clock started years ago. 109 11 min UiPath Isn't the New MuleSoft. It's the New Power Platform. UiPath's Maestro pivot has people asking if it's a MuleSoft competitor. Wrong question. Classic MuleSoft is integration plumbing at a different altitude; the real overlap is Power Automate and Copilot Studio. An architect on where UiPath, Microsoft, MuleSoft, and Salesforce actually compete, why Agent Fabric muddies the one seam that's genuinely contested, and how to lay out all four without two control planes fighting. 108 8 min The Agent Broke Prod at 2 A.M. and Left No Note You have written the gates, the sandboxes, the identity boundaries that stop an agent from breaking production. You never wrote the runbook for after it does. And you cannot run a blameless postmortem on an actor that emitted no trace, because your observability was built for deterministic services that fail the same way twice. 107 14 min Announced or Shippable? A Developer Reads the Stack Under Fiserv's agentOS Fiserv Forum will pitch agentOS as the headliner, riding a coherent stack: Finxact's temporal core, Data Compass on Snowflake, a stablecoin that treats dollars as just another ledger position. The architecture is real. An architect on the only question worth the flight, which layers you can actually touch today, and which are client-only pilots pitched into a market that wants proof. 106 8 min Nobody Budgeted for the Agent That Never Sleeps Per-seat SaaS budgeting assumed a human who logs off at six. An agent fleet meters continuously, and the cost model that governs it is a token economics your finance team has never seen. The meter was running the whole time, and the line item to catch it does not exist yet. 105 12 min Steal the Contract, Not the Catalog: Karpathy's LLM Wiki, Two Ways, for Regulated Banks Andrej Karpathy's LLM wiki went viral, but the same shape had already surfaced in Jake Van Clief's ICM and would land again in Google's Open Knowledge Format. One encodes workflow, one encodes trust, both are markdown in a folder. An architect on why the thing worth stealing isn't the tool, it's the contract, and how to run it on Azure or AWS infrastructure you already govern. 104 10 min The AI Employee Roster Was Built for a Plumber. Every Line Has a Bank-Sized Twin. The SMB pitch sells an AI receptionist to a plumber losing $500 a missed call. Read the same roster as a bank and every role maps to a line item a bank already measures. Here is the ROI, role by role, and the compliance tax that decides which ones actually clear the bar. 103 9 min Business Is the Programming Language In the AI era the constraint was never typing code, it's knowing what to build. Domain knowledge was always the differentiator; now, with shrinking teams and an AI-assisted SDLC, it's the whole job. An architect on why domain-blind developers ship the defects AI can't catch. 102 8 min Your RAG Assumes One Shape of Knowledge Putting the knowledge layer in plain text in git was the right call. Retrieval is where it quietly breaks, because knowledge arrives in at least four different shapes. A single vector store flattens all of them into the one it handles worst. 101 9 min The AI Factory Has Five Rungs. Most Banks Are on the Second. Shopify told staff to prove they can't use AI before asking for headcount. A bank CEO can't send that memo. Here's a five-rung maturity ladder for an AI software factory in a regulated engineering org, why each rung costs more inside a bank, and where the ceiling actually is. 100 11 min The Harness Went Multiplayer. Build What It Skips. You spent a year turning your personal agent harness into a moat. Then YC open-sourced a harness for the whole company. The layer moved up a floor, and the opportunity is not the harness itself. It is the review and accountability gate the fleet harness leaves open. 099 10 min Java's AI Moment Is Here. The Gap Was Never Technical. If your talent and business logic already live in Spring Boot, the reflex to rewrite AI features in Python is the expensive wrong turn. Josh Long is right: the fight is mostly integration, and that's the JVM's home turf. 098 16 min The Killer Feature Is the Signature Block launched Buzz, an open-source workspace where humans and agents work together, and everyone is calling it a Slack and GitHub killer. The killer part is not the part they are selling. Buzz gives every agent a cryptographic identity on day one, and for a bank that signature, not the consolidation, is the whole story. 097 14 min The Model Stopped Being the Product An open-weights research lab just posted a job that tells you where the money actually is. Nous Research, maker of Hermes, is hiring a Forward Deployed Engineer to install its agent inside enterprises. The role is the tell: capability has commoditized, deployment is the product, and for a bank the real risk is the capability cliff after the engineer rolls off. 096 9 min Only Thirty-Two Percent of the Time Can the Agent Prove It's Itself The agent identity debate is fought over which standard wins. The part nobody quotes is the empirical one: when you hand an LLM the cryptographic identity workflow the standards assume, it completes the job as little as 32 percent of the time. Agent identity is not just architecturally unsolved, it is broken at the execution layer. 095 9 min The Paperclip Maximizer Clocked In An OpenAI model broke out of its sandbox and hacked a real company to cheat a benchmark. The paperclip maximizer stopped being a thought experiment. For a bank running agentic AI, the gap between the goal you set and the constraints you forgot to write down is the whole risk model. 094 8 min The Customer Authorized a Human, Not a Fleet Open banking gave consumers the legal right to share their financial data, and the consent behind that right was designed for a person clicking allow for another person's app. Point an autonomous agent at that same button, let it hand off to a second agent, and the consent chain the rule was written to protect quietly snaps. The federation gap is not a payments problem. It is a consent problem, and it lands on consumer financial data first. 093 9 min When the Agent Reaches for the Card Your payment rails were designed around a human who pauses before they pay. Agentic commerce deletes the pause, and the authorization models racing to fill the gap do not agree on who is accountable for the charge. The design fork is not who moves the money. It is where the permission to move it lives, and whether that permission survives an agent-to-agent handoff. 092 12 min The Second Brain Doesn't Ship I argued markdown scales from a personal second brain all the way up to a company lattice, that the format never changes and only the graph grows. I was measuring one axis and missed a second. There are two axes of scale, not one, and the moment other people log in, plain files stop working. Here is where the second brain breaks, and the database-shaped thing that replaces it in production. 091 8 min The Bank Got Dumber While Its People Got Smarter Every engineer on your team is getting sharper with AI, and your institution is getting no smarter at all. The apprenticeship loop that taught skilled work for centuries moved into a private window, and nobody filed a ticket for it. 090 9 min You're Renting the Model. Own the Harness. Every quarter a lab you don't control ships a new model and your carefully tuned setup behaves differently. The model was always a rental. The harness around it is the only thing you own, and in a regulated shop it's the only layer you can actually audit. 089 7 min Documentation Is the Code Now I reached for docs-as-code because my only real constraint was time, not because I had a theory about it. The quiet satisfaction came later, watching a full team of architects adopt the same pattern on purpose and treat the documentation as the artifact they actually ship. 088 19 min The Model Was Never the Problem Fraud detection and BSA/AML get sold as model problems. They are data and integration problems wearing model costumes. Why the backbone, not the classifier, decides whether a financial-crime program works, and why the scarcest resource is the one person who can say yes. 087 7 min Streaming in Name Only You run the Event Hub. The vendor produces to it with a bare HTTP POST, gets a 201, and moves on. Next day the ETL batch is short and nothing on the write path can tell you what you lost. A technical look at the producer guarantees a POST throws away. 086 12 min The Field Guide I Wrote for My Family Beats Most Enterprise AI Strategies I built one agent field guide for a two-person law practice and a retail store. It turned out more disciplined than most 8-figure enterprise AI programs, because it starts with the paycheck, not the platform. 085 8 min The Load-Bearing Agreement Every team is judged by the fights you can see. The agreement nobody ever puts on the agenda is the one holding it up, or quietly bringing it down. 084 9 min The Gate Was Built for a Human Your security team is reviewing machine-generated code at human speed, and your developers trust output they never read. Both blind spots meet at a gate that has quietly become theater. 083 11 min We Inspect the Traffic, So We're Covered Three teams evaluate the same LLM proxy and each picks a different winner. Functionality and cost point at Azure API Management. The security lens points at a five-word sentence that is quietly false. 082 10 min An ERD Is the Floor, Not the Ceiling When the data team hands off a result set, sample rows, and a materialized view instead of a model, the solution team can't design anything. Why an entity model is the minimum shared language, and how to get it back. 081 10 min Foundry and Fabric, Not Foundry or Fabric I flew to Manhattan to learn whether Agent Fabric and Data 360 hand me a control plane I own or an elegant lease. The verdict on all four questions is that nobody owns the plane outright, and the most likely future is coexistence. 080 11 min The Architecture Is Clear. The Investment Isn't. Agent Fabric is a well-designed control plane for agent sprawl. The hard part of the investment case is not the license. It is the learning tax a new abstraction charges a team that cannot grow. 079 10 min Markdown Is the Knowledge Layer Karpathy quietly posted a gist for an LLM wiki. Two months later Google published the Open Knowledge Format, an open spec that formalizes exactly that pattern. The thing I argued the architecture corpus already was, plain text in git, is now the standard for all text-based knowledge, and it runs from your personal second brain all the way up to Miessler's lattice. 078 8 min The API Catalog Is the Tool Catalog After Build I argued the context layer agents read was already sitting in git: the C4 model, the ADRs, the corpus an architect keeps. Walking into MuleSoft Connect and the Data 360 workshop, here is the sequel. The tool layer agents act through is already sitting in the API catalog I govern. Here is what I am walking in to find out. 077 10 min Fifteen Pages, Nothing to Ship I drained a software developer retainer and got proof of time, not proof of work: meetings, design sessions, and a fifteen-page document no coding agent could build from. The only thing protecting my implementation partners is how slowly my side moves. I am dismantling that, and what I buy next has to ship. 076 9 min Your Size Was Always the Advantage In February I told my son that being small was an edge in the AI transition, not a liability. Four months and a pile of data later, the edge is real and it is perishable. Individual lawyers and their clients are moving faster than the firms are restructuring, and that gap is the whole risk. 075 8 min Architecture Is the Prompt I said the durable question after Build was whether the context layer lives somewhere I control or somewhere I rent. For an architect, that layer already exists: the C4 model, the ADRs, the DSL, plain text in git. One corpus, two readers, and governance lives in the diff. 074 8 min Agents Don't Commute The coding agents named at Build run inside the same WSL session as your developers, with their credentials, under sensors built for humans. The visibility stack sees hands, not minds. What the v2 golden image must add: MXC on, egress fenced, credentials short-lived. 073 8 min WSL Is Not Optional Every modern toolchain assumes Linux. Every bank-issued laptop runs Windows. At most regulated shops the answer is a jumphost, and the commute is quietly taxing your developers. How a small team makes WSL the paved road instead. 072 11 min Microsoft Conceded the Harness. Then It Moved the Moat. I went to Build to resolve a binary: control or governance. Microsoft shipped containment into the OS and made the harness bring-your-own, then moved the lock-in up to context. Open at the bottom, locked at the top, again. 071 8 min Governance or Containment: The Real Question I'm Bringing to Build The decisive axis at Build 2026 isn't Copilot Studio versus Foundry. It's whether Microsoft's governance can attach to a runtime I own, or only travels with its SDK and managed service. Here's the test I'm running and the decision it forces. 070 9 min Six Primitives for a Code Factory Ramp ships 30% of its PRs from a background agent. Stripe merges a thousand a week. Regulated shops have a small team and a change advisory board instead. Here are the six primitives required to run an autonomous code factory in a regulated bank, and why each one takes longer there. 069 9 min Who Scans Whom: The Agent Discovery Asymmetry Between MuleSoft and Microsoft MuleSoft's agent scanners catalog Microsoft Copilot and Foundry agents. Microsoft's Registry Sync catalogs Salesforce Agentforce. Neither catalogs the other, and that asymmetry tells you everything about where the registry war is heading. 068 9 min I Built Clerko: A Hermes-Powered Legal Agent for My Attorney Son How I used Hermes Agent by NousResearch to build a self-hosted AI assistant for my attorney son, with 28 Florida legal workflows codified and accessible via Telegram and email. 067 7 min My First Microsoft Build: What I'm Actually Watching The trip to Build 2026 isn't confirmed yet: last-minute decision, approval still pending. Here's what I'd be walking in to find out if I make it. 066 14 min What witness.ai Doesn't See witness.ai is a real product solving real problems. But in a Claude Code environment, the most dangerous attack surface sits below the network layer, and that's exactly where witness.ai has no visibility. 065 12 min Beyond the Rebrand: What Anypoint Omni Gateway Actually Changes Flex Gateway is now Omni Gateway. The runtime didn't change, but MCP/A2A native governance and Gateway Federation did. The more important story is what Omni Gateway reveals about Azure Foundry's Anthropic integration gap. 064 9 min The Bottleneck Was Never the Code Your engineering team uses AI every day. Your delivery velocity hasn't moved. That's the correct outcome for what you bought, and here's how to fix the decision, not the tool. 063 11 min When Your Team Starts Building Hand a team its first agent assignment and five engineers come back with five different builds. Here are the five patterns, and the distribution wall every one of them hits next. 062 9 min From Paperclip to Production: The Open-Source SDLC/SRE Bet Every Enterprise Is Weighing Paperclip AI and Hermes Agent are production-grade open-source platforms for enterprise SDLC and SRE automation. The question isn't whether to use them. It's when. And in regulated industries, that timing bet is harder than it looks. 061 9 min Claude Is Already in the Building Anthropic's Code with Claude 2026 shipped a Managed Agents platform. If you're wondering where it fits when Microsoft and Salesforce own your enterprise real estate, it already fits inside both of them. Here's the playbook. 060 12 min The Aircraft Carrier's Business Brain Daniel Miessler says most companies aren't ready for AI. Block.xyz says hierarchy is a 2,000-year-old information routing workaround AI makes obsolete. Both are right, and neither was written for a large regulated bank. Here's the enterprise translation. 059 11 min The Zero-Copy Promise Your core banking provider just gave you a Snowflake share. You plugged your ETL pipeline into it and things got worse. That's not a Snowflake problem; it's the wrong mental model applied to a new tool. 058 10 min Same Question, Different Worlds: What Your Team's AI Readiness Actually Looks Like You asked your engineering team to document a daily process. You got five different answers. That's not dysfunction; it's diagnostic data, and the most honest picture of where your AI transformation actually starts. 057 9 min I Use Microsoft Every Day. I Don't Trust Them. I've been told to change my dislike for Microsoft. What gets perceived as a dislike is actually an architectural discipline, one that applies to every enterprise giant, not just Microsoft. 056 12 min Who Issued the Agent? Your Okta federation governs human identity. Your Copilot Studio agents got Entra Agent IDs. These are different governance planes, and they don't talk to each other. 055 11 min The Agent Control Plane War: MuleSoft Agent Fabric vs. Microsoft Foundry + Agent 365 Microsoft and Salesforce are both pitching the same thing this month: govern every AI agent in your enterprise. If you're on Anypoint, you may already have one. 054 10 min Open at the Bottom, Locked at the Top: The Real Story of Microsoft Fabric Microsoft built Fabric on Delta Lake, Parquet, and open table formats. That's not generosity; it's a strategy. The lock-in didn't disappear. It moved. 053 10 min Claude Code: Where It Actually Runs Two enterprise paths for Claude Code: Azure AI Foundry or Anthropic Enterprise. For an Azure-invested architect at a regulated bank, the answer isn't what procurement assumes. 052 11 min TDX Day Two: The Floor Read Day Two on the TDX floor answered the keynote questions and raised a bigger one. MuleSoft MCP Bridge is GA, React is sandbox-only, Vibes 2.0 ships a CLI, and the real Salesforce decision was never technical. The final dispatch. 051 10 min Salesforce Headless 360: The Day One Read I walked into TDX expecting the maximalist Agentforce bundle pitch. Salesforce pitched the unbundle instead. Headless 360 decomposes Customer 360 into API endpoints, supports React and third-party agent harnesses, and deploys to ChatGPT, Gemini, and Claude. Here is the architecture read from Day 1. 050 10 min The Practical Value of a Salesforce Certification A classmate holds twenty-three Salesforce certifications. Another holds thirteen. I hold two architect-track certs and I keep asking the same question: what is any of this actually worth inside a corporate org where nobody asks about your Trailhead ranger status in the Monday standup? 049 8 min The Developer Experience You Can Bring Yourself Salesforce Agentforce Vibes blends copilot completion with vibe coding. Claude Code takes a different approach entirely. Here is what I found trying both before TDX, and which parts of the Salesforce dev AI bundle look optional. 048 8 min Salesforce Without the Full Buy-In Heading to TDX 2026 expecting the full Agentforce and Slack pitch. Here is what an Azure-invested architect might be able to opt into instead, and where MuleSoft, Agent Fabric, and Flex Gateway look like they earn their keep or don't. 047 11 min Agent as a Service Is Real. The Vendor Pitch Isn't. Gartner says 40% of agentic AI projects will fail. The builders see an infrastructure stack assembling: some layers solid, others barely sketched. Here's what AGaaS actually looks like when you strip the keynote slides. 046 9 min Claude Buddy Stopped Being a Tool and Became an Extension Four months ago I identified Claude Buddy's missing piece: learning, memory, identity. Today that integration ships. Claude Buddy v5 is a PAI extension, not a standalone tool. 045 8 min The Stack Nobody Designed Jensen Huang called OpenClaw the operating system for personal AI. Builder after builder arrived at the same architecture from below: memory, skills, heartbeat. The convergence is the signal. Here's what it means. 044 12 min Context Engineering Is Infrastructure, Not a Skill Fifty years of knowledge systems keep solving capture and failing at retrieval, or the reverse. The building blocks for context infrastructure are finally clear enough to define. Here's the blueprint. 043 11 min Same Person, Five Systems: The Data Identity Problem Banks Still Can't Solve Authentication identity and data identity are converging, and AI agents are the forcing function. Entity resolution, MDM, and why the 'golden record' keeps failing in banking. 042 10 min The Documentation Dividend: What AI Replacing Knowledge Workers Actually Looks Like Inside a Bank Daniel Miessler's expertise documentation gap, lattice architecture, and automation-first inversion, extrapolated to banking engineering with four concrete implementation patterns. 041 10 min Everything Is Skill Issue Karpathy says he hasn't typed code since December. The METR study says developers who use AI are 19% slower but believe they're 20% faster. That 39-point perception gap is the psychosis he's describing. 040 10 min The Sandbox Isn't the Hard Part NVIDIA's NemoClaw solves agent security with kernel-level sandboxing. The organizational governance that actually causes failures remains unsolved. 039 10 min Automating Chaos Produces Automated Chaos Sol Rashidi's 200+ deployments prove 70% of AI failure is organizational, not technical. The question is what you do with the execution surplus AI creates. 038 11 min AI Can Write Your Terraform. Can Your Organization Deploy It? AI-generated infrastructure code works. The question is whether your org has the GitOps maturity to use it, and where AI delivers without that prerequisite. 037 9 min Python Won't Run Your Bank: The Language Stack Reckoning for Enterprise AI Rod Johnson says stop letting data scientists architect enterprise AI. He's right, and the conversation he started deserves to go further. What TypeScript's rise tells us about where this is heading. 036 10 min Your Agent Is Not More Gas. It's the Thermostat. The discourse says AI adds cognitive capacity to the Four Burners stove. That's wrong. AI's real value isn't more gas. It's the feedback loop that detects drift and nudges you back toward desired state. 035 11 min The Architect's Dilemma: When Your AI Vendor Powers Both Sides Microsoft built Copilot Cowork on Claude. Your developers already run Claude Code. Same reasoning engine, two ecosystems, one governance gap nobody planned for. 034 11 min The Infrastructure Your Life Is Missing Jordan Peterson's seven life pillars aren't just 'more burners.' They decompose into two tiers: five application domains and two infrastructure services. The infrastructure tier is what the Four Burners theory never had. 033 10 min The Four Burners Are a Monolith The Four Burners theory says you must sacrifice parts of your life to succeed. But the 'cut a burner' conclusion isn't a law of nature; it's the predictable failure mode of a badly architected system. 032 10 min Dozens of Portals, One Identity: How an Architect Thinks About Enterprise Identity Consolidation When a financial institution operates dozens of customer-facing portals, each with its own identity system, consolidation isn't optional. OIDC, FAPI 2.0, FDX consent, and the question nobody wants to answer about AI agents. 031 12 min Architecture Without Architects: A Pragmatic Playbook for Regulated Banks You don't need a 40-person EA team to get every department on the same page about technology. Five well-maintained artifacts get you 80% of the alignment. Here's the playbook. 030 10 min Prompting Split Into Four Skills. Infrastructure Is the Fifth. Everyone's racing to learn four new prompting disciplines, but the real 10x gap isn't knowing the skills, it's encoding them as infrastructure that runs without you. 029 9 min The Reconciliation Loop: Desired State Is Not New. The Executor Is. Infrastructure engineers have run reconciliation loops for a decade. Daniel Miessler just described the same pattern with AI agents as the executor. 028 10 min The Personal AI Wars: A Developer's Guide OpenClaw vs. NanoClaw vs. Pi: who wins when the labs pull up the drawbridge? A developer's guide to the personal AI agent space and why model-agnostic infrastructure is the real bet. 027 7 min The Browser Is Not an API We solved this problem once with RPA and it failed. Now we're rebuilding the same duct tape with AI agents. The screen is not a programmable surface: it wasn't then, and it isn't now. 026 12 min Three Men, 160 Years, One Message A Victorian monk, a Jungian psychologist, and a modern entrepreneur converge on the same truth: radical responsibility, meaningful sacrifice, strength through action. 025 9 min The $100M Entrepreneur Who Teaches Discipline Like a Religion Alex Hormozi's philosophy: discipline beats motivation, execution beats ideas, and volume creates skill. Every single time. 024 11 min The Psychologist Who Made Meaning Mandatory Beneath the culture war noise is a Jungian clinician asking one question: if life is suffering, how do you make it meaningful enough to bear? 023 10 min The Victorian Monk Who Made Spirituality Dangerous Swami Vivekananda brought Vedanta to the West in 1893, but his real legacy isn't philosophy; it's the radical idea that weakness is the only sin and strength is the path to the divine. 022 9 min When the Regulation Hits Home: What a Government Developer Needs to Know About Staying Current in the AI Era I wrote one post for my son the lawyer and another for my daughter the CS student. Then I realized I owed my middle son, a government contractor writing code behind a regulation wall, the hardest conversation of all. 021 9 min When the Algorithm Hits Home: What a CS Junior Needs to Know About the AI Job Market My daughter read the post I wrote for her brother the lawyer and asked the harder question: what about the people building the tools? The honest answer for a CS junior facing 2027's job market. 020 7 min Judgment Is the New Moat In the agent era, execution is commodity. The only durable competitive advantage, for entrepreneurs and enterprises alike, is knowing which problems to solve. 019 9 min When the Plugin Hits Home: What Anthropic's Legal AI Means for Your Local Litigation Practice My son texted asking what 'restructuring your entire workflow' actually means for his litigation practice. The honest answer: the billable hour is under structural pressure, but judgment and accountability just became your most valuable assets. 018 8 min The 48-Hour Repricing: What a Legal Plugin Taught Banking About Its Software Future A 200-line Claude plugin erased $285 billion in software market cap. Banks weren't the target, but the shockwave reveals exactly where banking technology is headed next. 017 12 min When Everything Is Privileged, Nothing Is Banks classify every developer action as privileged. When 100% of your work needs a power account, dual accounts create friction without security, and modern JIT access fixes both. 016 8 min The 2026 Roadmap: From Ideation to Building Synthesizing ten posts into concrete building priorities. What I'm actually committing to build for regulated AI work in the next two quarters. 015 12 min Validation Hooks: The Technical Foundation for Higher Trust A deep-dive on implementing validation systems that enable climbing the trust ladder. Pre-execution gates, post-execution verification, and confidence scoring. 014 9 min OpenClaw and the Trust Ladder: What 145,000 Stars Tell Us About AI Autonomy OpenClaw validates massive demand for personal AI agents, but it's a case study in what happens when you skip straight to Level 4 autonomy without graduation infrastructure. 013 8 min The Trust Ladder: A Framework for Calibrating AI Autonomy Not all tasks deserve the same oversight. A 5-level model for matching AI autonomy to task risk, with graduation criteria for climbing the ladder. 012 8 min The AI Context Portability Problem Your AI infrastructure absorbs knowledge over time. Where that context lives, and who controls it, is an architecture decision with real consequences. 011 9 min The Agent Impersonation Problem: When AI Commits Code as You When Claude Code commits, it's 'you' in the audit log. Exploring the identity crisis at the heart of AI-assisted development. 010 8 min Non-Human Identities: When AI Agents Need Employee Badges AI agents acting autonomously need identity, just like human employees. Exploring what Non-Human Identity means for regulated environments. 009 6 min 2025: The Year of Loss and Clarity Some years test you. Some years teach you. 2025 did both. 008 10 min The Reference Architecture: A 4-Layer Approach to AI Tooling Moving beyond product recommendations to architectural thinking. A framework for organizing AI tools with appropriate boundaries and context flow controls. 007 9 min Your AI Toolkit is Probably Illegal A compliance audit framework for AI tools in regulated industries. Most workers are unknowingly creating liability. Here's how to assess your risk. 006 7 min The Compliance Tax: Why Your AI Productivity Isn't 10x (And Why That's Okay) Challenging the '10x productivity' marketing narrative. What regulated workers actually experience, and why 3x is still transformational. 005 8 min Claude Code vs Copilot CLI: Making the Enterprise Case GitHub is already approved. Copilot is already licensed. The benchmarks show statistical parity. Why would you invest in adding Claude Code? Here's the answer. 004 7 min The Clone Problem: Why Your AI Personal Agent Can't Follow You to Work What happens when you want a single AI that knows your whole life, but your employer operates in one of the most regulated industries on Earth? An architect's take on the hardest unsolved problem in personal AI. 003 10 min From Threads to Trust: A Framework for Measuring Your Agentic Engineering Growth How to measure and improve your effectiveness with AI coding agents using Thread-Based Engineering and the 90/10 simplification principle. 002 6 min What PAI's 7-Phase Loop Taught Me About Claude Buddy's Missing Piece Exploring how Daniel Miessler's Personal AI Infrastructure philosophy could enhance Claude Buddy's enterprise workflows, and what regulated environments can learn from personal productivity systems. 001 5 min Why I Built Claude Buddy: Bringing AI-Driven Development to Regulated Banking A journey from enterprise architecture skeptic to AI tooling builder, and what it takes to adopt AI-driven development in one of the most regulated industries.