Field Notes Blog

One control plane now watches ten Monument Labs products, keeps a falsifiable ledger for each, and proposes its own fixes through a gate it cannot talk its way past. Notes on running a portfolio on about ten minutes of attention a day.

An agent root-caused a caption bug in Camera Shy with a measurement harness, a four-arm experiment, and a debiased instrument. Total spend: about five dollars. The bug is the small story. The bill is the big one.

Part two: the details that separate a professional product from an obviously vibe-coded one. Link cards, finished metadata, social login, selection control, and layouts that behave.

Seven practices that make agentic development fast at Monument Labs: one stack learned deeply, a full product template, prototyping in real code, and why the prompt matters less than the machinery around it.

Daylight's autonomous analyst runs single investigations that cross data science, engineering, product, and support: from a metric anomaly to the exact user it hurt and the commit behind it. Three real cases, and a name for the discipline.

Roughly a fifth of Daylight's application code ships no features: it observes the business, keeps falsifiable beliefs with the SQL to re-check them, and escalates when reality drifts. Notes on what grew, and what it did to my job.

Ask software to turn someone's documents, evidence, and timeline into a written narrative and you inherit a specific risk: a confident sentence about something that never happened. Here is the design choice that lets anyone trace every line back to its source.

Two days at Meta. Thirty minutes now. How an operator and an agent collapsed a four-source weekly funnel review into one continuous session, and why DS directors should care.

Why every significant AI-built feature should ship with a verification artifact: an HTML page of screenshots, sample outputs, and SQL checks that lets a human verify the work in five minutes instead of fifty

Our content pipeline kept publishing posts that readers flagged as AI-generated. The grammar was clean, the facts were right, nobody cared. What it took to fix the voice, and a surprise about where 'more passes' stops helping.

When someone asks ChatGPT or Perplexity about your business, there's a good chance your website isn't part of the answer. Here's why, and what to do about it.

A real engagement where we used Claude Code to parse 1,271 co-parenting messages from PDF exports, structure them into searchable data, and generate a cited evidence report for discovery. What used to take 15 hours of paralegal work took 15 minutes.

Exploring how OpenClaw creates a persistent AI assistant experience using an event-driven Gateway architecture rather than a continuously running model.

A deep dive into what's actually happening under the hood of Claude Code: the ReAct loop, the tool system, the hooks, and why understanding this architecture matters more than using it

The unsexy, invisible AI implementations that actually drive ROI—and why the flashiest features are often the worst investments

Why the 'simple' traditional ML approach often costs 10x more than the 'overkill' LLM solution—and what it reveals about how we evaluate technology

Why organizations with refineries full of premium insights are watching decision-makers run on empty—and how to fix your broken insights pipeline

The uncomfortable truth about AI implementation and why starting with your boring problems is the smartest strategy

Why that 'quick' manual task is costing you more than you think—in money, morale, and missed opportunities

The universal struggle of having mountains of data but no clear answers—and what to do about it