Notes from the foundry
Engineering essays on generative AI, retrieval systems, and what it takes to ship intelligent software to production.
The 2026 Enterprise AI Build vs Buy Matrix: Platform Commoditization and Custom IP
Buying packaged vendor AI suites locks you into closed ecosystems with 4x markup. Building from scratch burns 18 months on commoditized plumbing. Here is the operational build vs buy decision matrix for enterprise CTOs.
Securing MCP Servers in Enterprise Networks With Zero-Trust Architecture
Model Context Protocol turns AI agents into active operators with system access. Here is how we design least-privilege auth, mutual TLS, and air-gapped sandboxing so enterprise security teams approve internal MCP deployments.
From AI Agent Pilot to Production in 90 Days, With the Gates in Between
Most agent pilots die between the demo and the deployment. This is the 90-day plan we use to get an agent into real production volume, including the four gates where the honest answer is sometimes to stop.
Synthetic Data & Automated Eval Pipelines: Beyond Naive LLM-as-a-Judge
Simple LLM-as-a-judge setups suffer from position bias, verbosity bias, and self-preference. Here is how to architect adversarial synthetic datasets, multi-judge consensus matrices, and deterministic assertion pipelines.
AI Agents Are About to Start Buying From You. Is Your Checkout Ready?
Agentic commerce protocols settled into a working stack this year: ACP for checkout, AP2 for payment authorisation, MCP and A2A underneath. Most merchant systems are built on assumptions that an agent breaks. Here is what to check before an agent tries to buy something.
The CFO's Guide to AI ROI: Measuring Process Automation Share Over Vanity Headcounts
88% of enterprise executives are raising AI budgets, but less than 20% can calculate their realized return on investment. Here is the operational framework for measuring straight-through processing rates, cycle-time compression, and unit cost reductions.
How We Build Synthetic Data Pipelines for Small Model Distillation
Relying on frontier API calls for high-frequency internal steps ruins unit economics. Here is how we build verifiable synthetic data filtering loops, distill multi-step reasoning into 8B open models, and drop inference bills by eighty percent without degrading accuracy.
MCP, A2A, and the Protocol Questions to Put in Your Next AI Contract
Agent interoperability standards settled faster than anyone expected. MCP for tools, A2A for delegation, both under Linux Foundation governance. Here is what that means for buyers, and the contract language that keeps you from being locked in.
Agent Adoption Is Not a Model Problem Anymore. It Is an Integration Problem.
Model capability stopped being the constraint. Roughly 46% of enterprises now name integration with existing systems as their biggest blocker to AI agent deployment. Here is what that work actually consists of, and how long each part really takes.
Who Manages the Agents? The Org Chart Question Nobody Answers Until Something Breaks
Agents are running in production at most enterprises and reporting to nobody. Here is the operating model that actually works: who owns an agent, who is on call for it, and what changes for the managers whose teams now include software that makes decisions.
Agentic AI Adoption in 2026: What the Numbers Actually Mean for Your Budget
79% adoption, 72% in production, 88% raising budgets, and 40% of projects predicted to be canceled. The agentic AI statistics for 2026 look contradictory until you notice what each survey is counting.
Your Agents Are Making Seats Redundant. Fix the Contract Before the Renewal.
Per-seat pricing is collapsing across enterprise software, and most buyers are still renewing on it. Here is how agent deployments change what you should be paying for, and the four clauses worth fighting for at your next renewal.
88% of Companies Running AI Agents Had a Security Incident. Ask the Blast Radius Question First.
Prompt injection is up sharply and most enterprises with agents in production have had an incident tied to one. The organisations that came out fine were not the ones with better prompts. They were the ones who had already decided what an agent could destroy.
Your AI Agents Outnumber Your Employees. Who Owns Their Credentials?
Non-human identities now outrun human ones by 45 to 1 or worse in most enterprises, and almost nobody can list their service accounts. Agent identity governance is the control that decides how bad your first agent incident gets.
Agent Washing: How to Tell a Real AI Agent From a Chatbot in a Vendor Demo
Gartner reckons only about 130 of the thousands of vendors selling agentic AI are the real thing. Here are the questions that separate an actual agent from a rebranded workflow tool, and what a truthful demo looks like.
Why 40% of Agentic AI Projects Get Canceled, and What the Survivors Did Differently
Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027. Having watched a number of them die up close, the causes are boringly consistent, and every one of them is visible in the first three weeks if you know what to look for.
The EU Just Gave You 16 More Months on AI Compliance. Do Not Spend Them.
The EU AI Act's high-risk deadline moved from 2 August 2026 to 2 December 2027. The transparency rules did not move. Here is what actually applies to your AI agents right now, and why the extension is a trap for anyone who treats it as free time.
The Effort Dial Is an Architecture Decision, Not a Setting
Claude Opus 5 ships with a low, medium, high effort toggle, and most teams set it once globally and forget it. In an agent loop, effort is a per-step decision, and treating it as a global default costs you money on the mechanical steps and reliability on the one step that mattered.
Three Tiers, One Loop: Routing Between Model Sizes Without Breaking Your Agent
GPT-5.6 shipped as Sol, Terra, and Luna on the same day. Tiered families make per-step model routing obvious in theory and expensive in practice. Here is what routing actually costs you in evals, caching, and debugging, and the two cases where it still pays.
Kimi K3 Is Open Weight. That Does Not Mean You Can Run It.
Moonshot released the weights for a 2.8 trillion parameter model under a Modified MIT license. Open weights make it legally yours, but a sparse MoE decouples compute from memory, and memory is what decides whether self-hosting is a real option for your team.