Twelve ways a multi-agent team broke, and the rule each one produced
Twelve real incidents from a five-person, five-agent engineering team, and the written rule each one produced.
6 min read ai-engineeringagentsprocess
Writing
Notes from the field.
Long-form essays on architecture, production AI systems, and how
engineering teams work when the assistants are part of the team. Each
one comes from something we built, measured, or broke.
Series · 5 essays
Leading AI-assisted teams
What changes when five engineers each drive their own AI coding assistant against one repository: coordination, decisions, and where the bottleneck moves.
Your status board costs 136K tokens a session
A coordination process that worked was quietly burning 136K tokens per session on every machine. How we measured it, and the ten rules that cut it.
5 min read ai-engineeringcontextprocess
The repo is the only shared state
With five engineers and five AI assistants, the repository is the only context they all share. What follows from that.
6 min read ai-engineeringteamsprocess
Rulings, not guesses: making "never assume" operational
Turning 'never assume' into a pipeline of questions, rulings and records that agents and humans both follow.
6 min read ai-engineeringprocessdecisions
The bottleneck is no longer coding
A five-point work package went from claimed to delivered in about fifty minutes. What that does to planning, review and the schedule.
5 min read ai-engineeringplanningteams
Series · 9 essays
Shipping LLM systems
Patterns from production LLM services: where to trust the model, where to take the decision away from it, and how to test what is left.
One proposal, one executor, one external write
Why the model should propose and code should execute: a state machine for external writes, with tamper detection.
6 min read llmagentsarchitecture
Never ask the LLM for the headline number
Let the model judge each criterion and let code compute the score. Fixing a match score users could not explain.
5 min read llmscoringarchitecture
Stop lying about what the rule statuses mean
A compliance panel that looked real and computed nothing, and how we replaced it with deterministic rules and an LLM fallback.
6 min read llmcompliancearchitecture
Vectors rank. Structured filters decide membership.
Cosine scores for right and wrong matches overlap across 0.20 to 0.62. Use vectors to rank and structured filters to decide.
6 min read llmretrievalpgvector
Three ways LLM JSON fails in production, and the guardrails that stuck
A Python dict rendered in the UI, truncated output, and the timeout cliff: three LLM JSON failures and the guardrails that held.
5 min read llmreliabilityparsing
Shipping an agentic rewrite behind flags that keep prompts byte-identical
How to rewrite a working assistant as a tool-calling agent behind flags that leave prompts byte-identical, using shadow mode.
6 min read llmmigrationtesting
Thresholds are config, not code
Why approval limits, tolerances and cut-offs live in a per-tenant config service, and the bugs that taught us to keep two read paths.
6 min read architectureconfigcompliance
Testing LLM systems without a cassette recorder
2,804 tests and zero recorded LLM calls: what mocking buys, what SQLite hid, and how to evaluate the model itself.
6 min read llmtestingquality
Disclosed limitations beat confident summaries
Design docs said inputs were sanitized and the code said otherwise. Why drift is normal, and the habit of writing limitations down.
4 min read llmdocumentationsecurity
Archive
Earlier writing
Architecture antipatterns
Common architecture antipatterns and how to avoid them
6 min read architecturesoftware_engineering
Software Architecture with Stability Patterns
Navigating the rough seas of software architecture with stability patterns.
3 min read patternssoftware architecture
Data Import/Export Using Firebase and MeiliSearch in React via Cloud Functions
How to build data import/export for your SaaS with Firebase Cloud Functions, and add fast full-text search with MeiliSearch.
2 min read reactfirebasemeilisearch
Typescript 5.0 release
A guide on new features in typescript 5.0
2 min read web developmenttypescript
Generative Adversarial Networks (GANs)
How Generative Adversarial Networks work: the generator, the discriminator, and the adversarial game between them.
7 min read machine_learningai
Error Boundaries in React
Learn how to gracefully handle errors in React applications using Error Boundaries.
1 min read reacterror-boundaryjavascript
An intro to React Context
When prop drilling gets painful: how React Context shares state across your component tree, and when to reach for it.
1 min read reactfirebasemeilisearch
Stability patterns
Explanation of common architecture patterns
4 min read architecturesoftware_engineering
Machine Learning Basics
The core ideas of machine learning explained from first principles: models, training, and how machines learn from data.
2 min read machine_learningai
AWS Relational Database Service
A practical introduction to Amazon RDS: what it gives you over self-managed databases, and how to get started.
3 min read awscoderds