Notes.
What we’re learning, what we’re shipping, what we’re refusing to do. Published on a rhythm, most of these posts come through our marketing-agent pipeline. Some land here first.
- 2026-10-26
RAG Evaluation: Score Retrieval and Answers Separately
RAG evaluation that works scores retrieval and the answer separately, so you can tell a search miss from a made-up answer. What to measure, and when.
7 min read · Engineering
- 2026-10-23
Tool Selection Quality: Is Your Agent Calling the Right Tool?
Tool selection quality: when to grade which tool an AI agent picked and with what arguments, and when to grade only the outcome it produced.
7 min read · AI Engineering
- 2026-10-21
Evals as CI Quality Gates for AI Agents: What Blocks a Merge
Evals as CI quality gates for AI agents: what to block on, where to set the pass rate, and what a gate costs per commit.
8 min read · AI Engineering
- 2026-10-19
Runtime Guardrails for AI: Check the Output Before Users Do
Runtime guardrails for AI check every output in production before anyone relies on it. What we run in our own products, and where a human still signs off.
8 min read · Engineering
- 2026-10-16
How Much Does a Custom AI Project Cost?
What drives AI project cost: scoping, data, evals, the production build and the monthly model bill, and why the pilot looks cheap but production does not.
8 min read · Build vs Buy
- 2026-10-14
AI Red Teaming for Business: Break It Before Customers Do
AI red teaming for business: what it is, how it differs from evals, which attacks to try first, and how a small team runs one before launch.
8 min read · AI Engineering
- 2026-10-12
AI Readiness Assessment: What to Check Before You Build
An AI readiness assessment checks data, workflows, decision owners and people before you build. What we check, what it should produce, when to skip it.
7 min read · AI Strategy
- 2026-10-09
AI Evals Are Not Software Tests
AI evals vs software tests: why a green test suite proves little about an AI feature, and why a public benchmark score is not an eval of your workflow.
8 min read · AI Engineering
- 2026-10-07
LLM as a Judge: When It Works and When It Lies
LLM as a judge: what the research says it gets right, the biases that fool it, and the checks we run before trusting a model to grade another model.
8 min read · AI Engineering
- 2026-10-06
Gemini 4 Argon for Business: Plan for the Swap
Gemini 4 Argon for business: Google announced it, but access is limited and the $2 and $10 price is introductory. What to do before it opens to you.
8 min read · AI Strategy
- 2026-10-02
When to Use AI Agents vs Simple LLM Calls
When to use agents vs simple LLM calls: a plain test for business teams. Most work is a fixed workflow with one model step. See where agents earn it.
8 min read · AI Strategy
- 2026-09-30
Scope Authorization for AI Agents, After GPT-6.1 Astra
OpenAI shelved GPT-6.1 Astra over scope and authorization. What scope authorization for AI agents means, why it lives in the harness, and how to build it.
8 min read · AI Strategy
- 2026-09-28
Claude Opus 5.5 for Business: The $20 Model Changes the Math
Claude Opus 5.5 for business: $4 and $20 per million tokens, thinking always on, cyber tasks rerouted. What changed, who gets it, and what to move first.
7 min read · AI Strategy
- 2026-09-25
The Tool Built to Publish Had No Rule About What Never Gets Published
An open source Claude Code skill turned your repo into a public launch video, with no rule keeping secrets out. Here is the fix we got merged, and the lesson.
8 min read · AI Engineering
- 2026-09-23
AI for Solo Practitioners: The Setup Before the Tools
AI for solo practitioners: 71% of solo lawyers use it and 57% have no policy. The three decisions to write down first, and what stays out of the chat window.
8 min read · AI for Law Firms
- 2026-09-21
Common Claude Mistakes Beginners Make (and the Fixes)
The common Claude mistakes beginners make start with expecting one prompt to do it. Anthropic's data shows 13 rounds for a typical article. The fixes.
7 min read · Learn Claude
- 2026-09-18
How Often Should Your Business Switch AI Models?
How often to switch AI models: Google shipped three Flash models in six weeks. Here is the upgrade policy that makes a release a choice, not a fire drill.
7 min read · AI Strategy
- 2026-09-16
Two Years of Building With AI, in My Own Usage Logs
Two years of building with AI, measured in my own usage logs: 6.5 billion tokens in Cursor, 56.8 billion in Claude Code, and what changed in how I work.
7 min read · AI Engineering
- 2026-09-14
OpenAI Agents API: The Codex Harness Is Now for Rent
The OpenAI Agents API is in public beta: the Codex harness as a managed service. What it runs, what it costs, who cannot use it yet, and what stays yours.
8 min read · AI Strategy
- 2026-09-11
Claude Fable 5.1 for Business: What Changed, Who Gets It
Claude Fable 5.1 for business: what changed from Opus 5, what it costs, who can get Mythos 5.1, and the data retention line that decides whether you can use it.
8 min read · AI Strategy
- 2026-09-09
AI Consent for Patients: What We Built Into a Dental Notes Tool
AI consent for patients is an architecture decision, not a form. Shipping AI clinical notes: what gets recorded, who hears it, what the model may not say.
8 min read · AI Integration
- 2026-09-07
GPT-6 Astra for Business: What Changed and What Did Not
GPT-6 Astra for business: same list price as Claude Fable 5.1, a 1M context window, and phased access. What changes for an operator, and what does not.
7 min read · AI Strategy
- 2026-09-04
Long-Running Agents in Production Just Got Cheaper
Long-running agents in production used to be priced out by context. Cache reads at 2.5% of input change the math. What it fixes, and what it does not.
7 min read · AI Engineering
- 2026-09-02
MCP Tool Bloat: Why Agents Get Worse After Integrations
MCP tool bloat is why agents degrade after you connect more servers. Here is what tool definitions cost in context, and the three fixes that actually work.
7 min read · AI Engineering
- 2026-08-31
Claudeforce: What Salesforce in Claude Means for SMBs
Claudeforce puts Salesforce inside Claude with 37 prebuilt sales skills. Here is the real signal for any business whose systems are not yet agent-ready.
7 min read · AI Integration
- 2026-08-28
Common AI Integration Failures: Why Projects Stall
Common AI integration failures are not model failures. Here is where business AI projects actually stall, and the five things we check before we build.
6 min read · AI Integration
- 2026-08-26
Claude Skills for Business Workflows: What Changed
Anthropic moved the Skills API out of beta on August 19, 2026. What a Claude Skill really is, when it beats an MCP server, and why a business should care.
7 min read · AI Engineering
- 2026-08-24
Provider-Agnostic Agents After MCP and A2A
Building AI provider-agnostic agents got easier now that MCP and A2A sit under one neutral foundation. Here is the layer that still locks you in today.
7 min read · AI Engineering
- 2026-08-21
Automating the Software Development Lifecycle With AI: Where It Pays and Where It Does Not
Coding was never the bottleneck. A map of the software development lifecycle, where AI automation actually compounds, and where we deliberately keep humans.
7 min read · AI Engineering
- 2026-08-19
AI Code Review Automation: What to Automate, and What to Keep Human
AI code review automation fails when you point a model at a diff and ask if it looks good. Here is the layered review setup we actually run, and why it holds.
7 min read · Field notes
- 2026-08-12
Using AI Agents for Large Migrations and Codemods
A 400 file migration is the best first job to hand a coding agent and the worst one to run unsupervised. Here is the pattern that makes it safe.
7 min read · AI Engineering
- 2026-08-11
The Model Org Chart: Staffing AI Coding Agents by Role
Stop asking which AI model is best. How I staff coding agents by role: planner, PM, parallel workers, and a rival-vendor reviewer via codex-plugin-cc.
7 min read · AI Engineering
- 2026-08-10
What People Actually Use Claude For, and It Is Not Coding
Anthropic studied 1.2M Claude Cowork sessions. Coding was 8.7%. Operations and writing were half. Here is what that means for how you learn.
7 min read · Learn Claude
- 2026-08-07
Multi-Agent Systems Fail at the Merge, Not the Fan-Out
Parallel subagents rarely break while running. They break when results come back together. Why the merge is the weak point, and how to harden it.
7 min read · AI Engineering
- 2026-08-05
Agent Observability: What to Trace When Nobody Is Watching
Agents fail quietly. Here is what to trace, which OpenTelemetry GenAI spans matter, and how to tell a broken agent from an expensive one.
7 min read · AI Engineering
- 2026-08-03
Forward Deployed Engineers: When the AI Lab Sells the Build
The model labs now sell the integration work too. What that structural change means for how you evaluate an AI vendor, including us.
6 min read · AI Strategy
- 2026-07-31
Claude Opus 5 for Business: Which Tier to Actually Run
Anthropic shipped Claude Opus 5 on July 24, 2026. Here is how I pick between Opus, Sonnet, and Haiku for real business work without overpaying.
8 min read · AI Tools
- 2026-07-29
MCP Servers in 2026: What the New Stateless Spec Changes
The MCP spec finalized July 28, 2026 with a stateless core, extensions, and tighter auth. What an MCP server is and what actually changes for your agents.
7 min read · Building AI Agents
- 2026-07-24
AI Agents Are Now a Metered Bill. How to Budget for Them.
AI agents moved to metered, per-task billing in 2026. Here is how to budget for agentic work and keep the bill from surprising your business.
7 min read · AI Strategy
- 2026-07-22
Human-in-the-Loop Agent Design: Where to Put the Person
The four places a person can sit in an agent workflow, how to pick one per action, and why the review-everything approach fails first.
7 min read · Building With AI
- 2026-07-20
Eval-Driven Agent Development: The Habit That Keeps Agents Alive
Most AI agent pilots never reach production. Eval-driven agent development is the single habit that separates the ones that survive a year.
7 min read · AI Engineering
- 2026-07-15
Your AI Project Is Really a Data Project
Data readiness for AI is why most projects fail before the model. What AI-ready data means, and how to tell if yours is, before you build.
7 min read · AI Strategy
- 2026-07-13
Guardrails for AI-Generated Code: The Hard Part Is No Longer Writing It
Coding models now clear 80 percent on SWE-Bench. Writing code is solved, trusting it is not. Here are the guardrails we put on AI-generated code.
7 min read · AI Engineering
- 2026-07-10
Claude vs ChatGPT for Getting Real Work Done: How to Pick (and When to Use Both)
Claude vs ChatGPT for getting real work done. A plain-English, multi-provider guide to picking the right tool for the job, with a simple way to decide.
6 min read · Learn Claude
- 2026-07-08
Agentic Commerce Just Went Live at the Card Networks. Here Is What It Means for a Retail Business.
Agentic commerce is live at Visa, Mastercard, and OpenAI. A plain-English guide to what AI shopping agents mean for a retail business, and where to start.
6 min read · AI Integration
- 2026-07-06
OpenAI Just Split GPT-5.6 Into Three Tiers. Here Is Which One Your Business Should Actually Run.
OpenAI split GPT-5.6 into Sol, Terra, and Luna. Here is a plain-English guide to which tier fits which business job, and why the cheaper one usually wins.
7 min read · AI News
- 2026-07-03
Gemini 3.5 Flash Just Got Agentic. Here Is What the Cheap Fast Tier Now Unlocks for Your Business
Google's Gemini 3.5 Flash brought computer use and agentic action to the cheap, fast AI tier. Here is what that unlocks for a business, and when to use it.
7 min read · AI Integration
- 2026-07-01
HIPAA-Compliant AI Is a Plumbing Problem, Not a Model Problem: What We Learned Shipping It on Vertex
HIPAA-compliant AI integration is mostly a data-plumbing problem, not a model choice. What we learned shipping a HIPAA-grade AI feature on Vertex AI under a BAA.
6 min read · AI Integration
- 2026-06-29
Prebuilt AI Agents Can Now Run Your Back Office. Here Is the Build-vs-Buy Line in 2026.
Anthropic shipped Claude for Small Business: 15 ready-to-run agentic workflows. Here is when a prebuilt back-office agent is the right call, and when you build.
6 min read · AI Integration
- 2026-06-26
Your First Week With Claude: What to Actually Do, and What to Skip
A plain-English first week with Claude for people who want real work done, not a tour. What to actually do, what to skip, and the beginner mistakes to avoid.
7 min read · Learn Claude
- 2026-06-24
Every Lab Just Shipped a No-Code Agent Builder. Here Is When I Actually Use One (and When I Still Build)
Google, OpenAI, and Anthropic all shipped no-code agent builders. Here is the honest build-vs-buy line: when a no-code agent is enough, and when it is not.
7 min read · AI Strategy
- 2026-06-22
Every Major Lab Just Shipped an Agent Control Plane. Here Is What It Means for Your Business
OpenAI, Google, and Anthropic all shipped agent control planes. What an orchestration layer does, when a business needs one, and how to stay provider-agnostic.
7 min read · AI Strategy
- 2026-06-19
Context Engineering Is the Real Job: How I Decide What an Agent Gets to See
Context engineering, not prompt wording, separates an agent that demos from one that survives production. Here is how I keep my own agents lean and reliable.
7 min read · AI Engineering
- 2026-06-17
AI Document Review Is the Wedge for Legal AI. Here Is Where I'd Actually Start a Firm.
AI document review for law firms in 2026: why it is the wedge, the assistant to agent shift, and where a small firm should actually start. An operator's view.
7 min read · AI for Law Firms
- 2026-06-15
Agentic Coding in Production: What Actually Changes When Coding Agents Move From Demo to Daily Driver
Agentic coding in production is not autocomplete. Here is how I delegate real dev work to coding agents, what I gate, and why usage-based billing now matters.
7 min read · AI Engineering
- 2026-06-12
Hundreds of Parallel Subagents Just Landed in Claude Code. Here Is When I Actually Reach for Them.
Anthropic shipped Dynamic Workflows and a 1M-token context with Opus 4.8. Here is when fanning out to parallel subagents helps, and when it burns tokens.
7 min read · AI Engineering
- 2026-06-10
Buying AI One Year at a Time: How I Keep Vendor Lock-In Out of What I Ship
Companies are buying AI one year at a time. Here is how I think about AI provider lock-in risk, and how I architect so swapping a model is a config change.
7 min read · AI Strategy
- 2026-06-05
Reliability Is a Harness Problem, Not Just a Model Problem
The smartest model still ships broken work if the harness around it is weak. Here is what I do to keep coding agents reliable once they hit production.
7 min read · AI Engineering
- 2026-06-03
Retail AI in 2026: What Actually Ships to the Floor, and What Stays a Demo
Retail AI is moving from pilots to the sales floor in 2026. I break down what ships into store operations, what stays a demo, and how to tell them apart.
7 min read · AI Integration
- 2026-06-01
Claude Sub-Agents, in Plain English: When Multi-Agent Architecture Earns Its Keep
Claude sub-agents architecture sounds like a developer problem. For operators, the real question is when multiple agents beat one. Here is the honest frame.
5 min read · Anthropic Ecosystem
- 2026-05-29
Claude API vs Hosted AI Tool: When Operators Should Build Their Own
Claude API vs hosted AI tool is a build-or-buy call, not a tech debate. A field guide for SMB and mid-market operators deciding when custom beats subscribing.
10 min read · Build vs Buy
- 2026-05-27
OpenAI Operator and the rise of browser agents, when computer use is the right tool for your operations team
OpenAI Operator, Anthropic Computer Use, and Project Astra let AI drive a browser. When that matters for a business, and when a plain API is faster and cheaper.
6 min read · Field notes
- 2026-05-27
AI for law firms, what actually ships and what stays a pilot
An operator's view of AI for law firms in 2026: where the wins are real, where the risk concentrates, and how a firm should sequence its first six months.
8 min read · Vertical
- 2026-05-26
Gemini 3 Pro for HIPAA-grade AI, what we shipped on Vertex AI for Smile PreVue
We picked Google Vertex AI Gemini 3 Pro for a HIPAA-grade dental AI product. Here is the build, the compliance work behind it, and what we would do differently.
6 min read · Field notes
- 2026-05-25
How to integrate AI into a business workflow without burning your first six months
How to integrate AI into a business workflow: what the work actually looks like, where teams get stuck, and the thin-slice path that ships.
8 min read · Methodology
- 2026-05-25
Hosted agents across Anthropic, Google, and OpenAI, which one for your business
Three labs, three hosted agent platforms, and a real decision framework. How we pick across Anthropic, Google Vertex AI, and OpenAI Assistants for production work.
6 min read · Field notes
- 2026-05-24
MCP in plain English, connecting your existing tools to Claude without rebuilding anything
Anthropic's Model Context Protocol lets Claude talk to your existing software. What that means for a business that isn't a dev shop, and what to check first.
6 min read · Field notes
- 2026-05-23
Claude Code for operations teams, yes, the same tool the engineers use
Claude Code is marketed at engineers, but the agent loop underneath it is general-purpose. Here is how we deploy it as an ops runtime for non-engineering teams.
8 min read · Field notes
- 2026-05-23
Anthropic's hosted agents, what they actually are, and when to use them in your business
Anthropic now runs the agent runtime for you. What it means in plain English, where it fits a business, and where you still want a custom build.
6 min read · Field notes
- 2026-05-22
Figure it out, build it, ship it, how we approach every engagement
Our method for shipping AI and software inside real businesses. Borrowed from a 200-page evaluation that seeded a spun-out QA/QC product.
3 min read · Field notes
