Written for the person who has to sign off on it, not for the person who has to build it. No vendor pitch, no benchmark charts, and the failure numbers included.
At the TinyGo hack session in Seattle I flew a Parrot minidrone through a stunt routine from a Go program on my laptop, on a library written by one of TinyGo's maintainers, who was in the room. What the code does, what broke, and what the rest of the conference taught me about where Go is going.
ComplianceKit was a working product for licensed daycares in three states: rules engine, document chasing, mock inspections and its own PDF e-signature, in one Go binary. Here is why I built it, why Go, the timeline, and why the hard part turned out to be everything that is not code.
Three Go services I wrote and run, read back from their own source: a childcare compliance backend with its own PDF e-signature, a first-party analytics collector that has taken 180,000 events in five megabytes of memory, and the public copy of that collector. What Go made cheap, where I cut a corner I can now point to, and what I would change.
We run a paying product, a staging copy of it, about fifty backend demos and ninety-six static sites on two of the smallest servers Hetzner rents plus Cloudflare's free tier. The trick is a placement rule, a memory cap on every container, and a coding agent that reads the inventory before it touches anything. Here is the whole method, including the part that is currently on fire.
We stopped renting analytics and wrote a collector that runs on our own domain, enriches every event at the Cloudflare edge and stores raw rows in SQLite on a server we control. Having every row turned out to matter less for the dashboards than for the corrections: proxy networks posing as companies, mail scanners posing as readers, crawlers arriving by the back door. You can only fix what you can re-read.
An application performance product sells three things: collection, storage and a reason for a person to look. On a few small servers the first two already exist, for free, in the reverse proxy's request log and in /proc. What is scarce is the look. We gave that job to a script and a coding agent, and its first run found two problems we did not know we had.
Rails World 2026 opened with DHH retiring the pencil and closed with Aaron Patterson explaining the C++ "as-if rule". The internet scored it as a feud. It was a narrower and more useful disagreement, about who keeps the promise that the code does what it says, and a working Rails shop has to pick a side of it by Monday.
TypeSafe's Jev puts a confidence number on every decision it makes, and the argument about whether that number can be trusted is the most useful AI argument of the month. A model that says how sure it is turns adoption from "is it good enough?" into two questions a manager can actually own: where does the line go, and who works the queue below it.
The creator of Rails flew a thousand Rails developers to Austin to tell them he no longer writes code, his flagship product is moving to Rust and native apps, and the correct emotional posture is total optimism. A review of the most confident hour in framework history, and the one question it never got round to.
TypeSafe's Jev returns typed decisions with probabilities instead of text, in a few hundred milliseconds, for a fraction of a cent. The product may or may not last. The split it exposes will: models are starting to specialise by the shape of what they return, and a judgement that costs 100 ms can sit inside a hot path the way a regex does.
TypeSafe's Jev prices a typed machine decision at a fraction of a cent, and within a week the open clones and the fast-follow arguments had arrived. If the model layer commoditises that quickly, the value moves to three things a vendor cannot ship you: the definition of the decision, the labelled record of how past decisions turned out, and whoever answers for the consequences.
Postmortems are very good at making a failure make sense, and a failure that makes sense feels handled. Now that a model can write a flawless root-cause section in ninety seconds, explanations are free and change is as expensive as ever. The only part of the document that matters is the part nobody reads.
Inference prices fell by roughly half again this week and the "tokens too cheap to meter" argument is winning on the numbers. The price was never the reason to keep deterministic tools. The expensive thing in a production system is a different answer on the second run, and cheap tokens make that thing more common, not less.
Claude Code started reading AGENTS.md this week, which means one file is now consulted by every coding agent on every task, hundreds of times a day, more than any wiki page has been read in the history of your organisation. Most of them are forty lines of contradictions written in an afternoon.
A senior security engineer wrote this week that he spends three quarters of his time on AI and it has robbed him of his enjoyment of the work. He is right about the symptom and wrong about the cause. The value from these tools is landing in the least glamorous places in the company, and the people looking at the frontier are looking the wrong way.
A nine-author study this week pulled a coding agent apart into its components and measured each one across 176 configurations. The findings are less exciting than any vendor slide and more useful than all of them, and they hand the buyer four questions no benchmark answers.
The anxiety of the last month has been about falling behind. The actual career risk sits in the other direction: the moment a model's output carries your signature. The military found that out this week. Law firms are about to.
Three of the most-read essays on Hacker News this week were about the same thing: writing with a model makes you sound like everyone else who writes with a model, and everyone can tell. Here is the mechanism, the list of tells, and the one way to use the tool that does not cost you your voice.
A coding agent offered free this month was found snapshotting workspaces, git history included, to the vendor's cloud. It is the second such story this year. The tool you let touch your repository has more access than any contractor you have ever hired, and nobody ran a background check.
A programming language shipped this week with a 99 percent AI-written compiler and no mention of the forty-year-old field it reinvents. The failure was not generation quality. It was that building got cheaper than reading, and nobody put a gate between them.
The model leaderboard reshuffles every six weeks and everything you learn about a specific model has a half-life of months. Five things do not decay, and they are the same five that the people running the swarms say they still need humans for. This is a list for the engineer who is tired.
Every team that bought coding agents this year got the same result: the generation side of the shop sped up ten times and the shipping rate barely moved. The constraint moved to the one seat nobody re-engineered, and it is the seat with a person in it.
Two mathematicians spent a year on a problem, a rumour of their approach reached a lab, and a swarm of agents reportedly reproduced it in days. Terence Tao's warning about open science applies just as well to your product. What was defensible last year is now a prompt.
There are dozens of well-funded companies building the same AI agent for the same back office, and they are not wrong about the capability. They are aimed at the wrong scarce resource. The constraint was never labour, and it was never the model.
Not the ones in the headlines. The real order is set by volume and reversibility, and it is remarkably consistent across every industry we walk into.
Everybody publishes the accuracy number. Almost nobody publishes the failures. The failures are the entire product, because they tell you exactly which gate to build.
Everyone assumes a review gate is a tax on throughput. Measured against the task it replaces, it is the cheapest thing in the building — but only if you build the interface properly.
Not answering hard questions. Not calming angry customers. Typing. Here is how to get the actual number for your own company in about three days, before you spend a dollar on AI.
A surprising share of what gets logged as a model failure is the system faithfully reporting a conflict that was already in the company. Grounding does not just reduce errors — it audits you.
Service businesses obsess over lead cost and ignore the leads they already paid for and then dropped. The arithmetic is brutal and almost nobody has run it.
Any team that has actually run this in production has a list of what the system got wrong, sorted by category, with counts. Asking for it takes ten seconds and separates the operators from the demos.
Accuracy is the wrong headline number. The number that decides whether a system is deployable is how well it knows what it does not know.
Different industries, different words, same fifteen jobs. If you recognise more than eight of these, you have a straightforward automation project and you do not need anyone to tell you that.
The model is a commodity that gets better every quarter without you. The reviewable action pipeline is the thing that is actually hard, actually yours, and actually worth money.
The model is a commodity that improves on somebody else's budget. The thing your competitors cannot copy is the twenty years of operational data sitting in the system you complain about.
They do not fail for technical reasons. They fail in five specific, predictable, entirely preventable ways, and you can check for every one of them before you sign anything.
Support organisations feel infinitely varied from the inside. From the outside they are a very short list, repeated, with different names attached.
It works brilliantly for one person. It falls apart the moment it becomes a process. Here is exactly where the seam is, and why it is not an argument against using it.
The thing that gets an operations director, a compliance officer and a nervous CFO to say yes is not the model. It is being able to answer, in eleven months, what happened and why.
None of them are technical. All of them are answerable by anyone who has actually deployed this in a real company. Print them out and bring them to the demo.
We build a great deal in n8n and recommend it constantly. It is also the wrong tool for about a third of what people put in it, and the failure is always the same one.
Not a prompt. Eight structural controls, none of which depend on the model behaving well, and all of which you can build before writing a single line of agent code.
If after reading it you want somebody to build the thing, that is what we do.