Field notes from the third way.
What we learn shipping software with AI agents, automating operations, and governing AI in real companies — written by the people doing it.
We build software with AI agents for a living, so this is where we publish what that actually means: the cost models we run with clients, the autonomy rules our engineers work under, and the security findings from real assessments. No hype, no doom — just field notes from the middle ground between "AI changes everything" and "AI changes nothing".
Looking for something practical first? Start with our services or see the model in action in our case studies.
We publish when we have something worth saying — usually two or three times a quarter — rather than on a content calendar. Every post is written by the person who did the work it describes, and every number in it comes from a real engagement. If a topic here raises a question about your own situation, the fastest answer is a conversation.
A free Beginner course on Claude Code — the habits, not the commands
Every Claude Code tutorial teaches the commands, and the commands change every month. Our free Beginner course — five lessons, a capstone, three electives — teaches the habits that survive the updates.
Read the post →Everyone uses AI. Almost nobody gets returns.
Two 2026 research reports arrive at the same paradox from opposite directions. Adoption is nearly universal, returns are vanishingly rare — and the minority that gets them changed the work, not the shopping list.
Read the post →Agents don’t need memory, they need documentation
A developer who spent a year building agent-memory tooling concluded the category is flawed: agents don’t need recall of past chats, they need documents they can read and you can audit. Here is what that means for a firm that writes no code.
Read the post →What the third way actually is
“AI enablement” has become a bucket everyone sells into. We are not bidding for that word. The third way is a delivery model with four tests — and every one of them is a mechanic our delivery board enforces, not a value we aspire to.
Read the post →What your always-on agent wakes you for
The launch post says the quiet part out loud: the agent “comes back when it needs a decision.” That sentence is the whole design problem.
Read the post →Your 10-page system prompt is a smell
The best models now reason from context. Most instruction bloat is information that belongs in retrieval, memory, or tools — not prose.
Read the post →Nine famous websites, zero agent-ready
Stripe scored 35. The New York Times 44. Anthropic — the company that makes Claude — 30. Here’s what the numbers actually mean, and how to read your own.
Read the post →Your next customer might be an AI agent
An insurance executive says a shopping agent strips the legal details out of quotes. Most brands don’t even know they’re being visited.
Read the post →Falling token prices won’t fix your AI budget
The unit price of machine-generated work keeps falling. The expensive parts of using AI were never the generation.
Read the post →Your law firm could have its own Claude plugin
Claude now runs books and chases invoices for small businesses — but not for law firms. We build the missing version: one that knows your matters, drafts your routine work, and sends nothing without a lawyer’s OK.
Read the post →Claude Architect Foundations (CCAR-F) exam prep
Anthropic's "Foundations" architect exam is the hardest of the three Claude certifications — more technical than the Developer one. I passed it; here is the prep plan I would follow, down to the only mock exams at real exam difficulty.
Read the post →Claude Architect Professional (CCAR-P) exam prep
I passed Anthropic's Architect Professional exam with zero preparation, booked three hours beforehand. Not a flex — a statement about what the exam tests: org-level design judgment, governance, and knowing where AI fits.
Read the post →Claude Developer Foundations (CCDV-F) exam prep
Of Anthropic's three Claude certifications, the Developer one is the fair one — if you already build with the API, you are most of the way there. Here is the one-to-two-week plan I wrote for a friend who asked for the fastest path.
Read the post →Jev and typesafe AI: judgment, not text
Most "AI feature" plans end at the same place: prompt a chatbot, ask for JSON, and write a parser robust enough to survive the days it doesn’t. Jev — TypeSafe AI’s first System One model — does something different: it returns typed judgments and probabilities your code can branch on.
Read the post →Shadow AI: the security risk already inside your company
In every assessment we run, we find AI tools the security team did not know existed. Shadow AI is not a future risk — it is a present, unaudited data exfiltration channel.
Read the post →What we actually let AI agents do (and what we never will)
"AI-assisted development" is uselessly vague. Here is our actual autonomy policy: the exact lines we draw, and why they sit where they do.
Read the post →Human-only vs. AI-only vs. the third way: a cost model
Everyone has an opinion on AI and software teams. Fewer people have run the numbers. Here is the cost model we use with clients, derived from real engagements.
Read the post →