Charity Majors on trust accounting for AI-written code
- YT :: https://www.youtube.com/watch?v=HC8T1OlgYi0
- Original title :: Stop being skeptical about AI for development with Charity Majors
Charity Majors (CTO, Honeycomb) was an AI skeptic in 2025 and changed her mind in late 2025 when Opus 4.5 landed — though she argues the real shift was the harnesses, not the models. The core idea: shipping code nobody read is a when, not an if, and the honest framing is a trust account. If you debit trust at code creation, you have to credit it back somewhere else — tests, evals, conformance testing, telemetry, guard rails.
Reading code you did not write
The useful question is not "would you ship unread code" but "what would it take". Her answer borrows from ops and QA, two disciplines software engineering has been snobbish about, which have always validated software written by unreliable agents (i.e. other humans). You already have not read most of your codebase.
She quotes Chad Fowler's Phoenix-architecture line — immutable infra, stateless services, blue-green deploys all share the premise "never fix a running thing, replace it" — and extends it to application code: when generation is cheap, editing in place accumulates entropy while replacement resets it. Code as cattle, not pets. She is explicit about the limit: anyone who has done a hard database migration should be humble about how far this goes.
Nondeterministic systems demand more engineering discipline, not less. Predict-the-failure unit tests are a low bar; the missing layer is conformance, smoke, load and fuzz testing — "how do I know this performs within the boundaries of the last code I generated" without reading it.
Code review is overloaded
"Do we want this in our product", API design and architecture are great human conversations. Reading for syntax and bugs is a teaching opportunity at best and shouldn't be the main event. She points to Intercom/Fin as the high-water mark: a pre-AI company with real discipline that went AI-native, ships a Ruby monolith hundreds of times a day, and runs AI-validated PRs carrying the wisdom of their most senior engineers on every diff.
The two camps
Neither side is making it up. Enthusiasts see real wins and a real race; the on-call crowd sees melting mental models, slop, and reliability going backwards. The split usually falls exactly along who carries the pager. Gergely cites Meta's recent flurry of SEV0s — concentrated in Instagram and WhatsApp, where reliability staff were cut — and says engineers at other VC-funded and public companies quietly report the same. Intercom published the honest version: 18 months of declining reliability and code quality, only recently starting to recover.
Her ask: tell the whole story. Publish the wins with the costs, or the two camps keep assuming each other are liars.
Norms at Honeycomb
"There is no human in the loop — you own the loop." No hiding behind what Claude said. Baseline etiquette: never send someone something you haven't read; if it takes them longer to read than it took you to generate, it's slop and it's disrespectful. AI as a shortcut to avoid thinking versus AI to think more rigorously — for core job functions, the second one.
Software is the killer app for AI because software is logic and language, and unlike legal briefs it can be validated: compiled, tested, checked. If a workflow can't be automated in software, other industries have no chance. Notably she finds AI worse at prose than at code — writing is thinking on paper, and she has stopped trying to shortcut it.
Observability
Twenty years of DevOps was one attempt to build a feedback loop from people writing code to that code in production, and it failed. The dev/platform split itself is a healthy separation of concerns; the missing piece is fast feedback.
Metrics and logs are system exhaust from third-party software you don't own — store them somewhere cheap. Your own code deserves telemetry as a product decision: one wide event with all connective tissue, because the value of a wide event grows combinatorially — the 30th field on a 29-field span is worth more than all the others.
Auto-instrumentation with OpenTelemetry is now good enough that instrumenting is faster than not instrumenting, and the models are trained on the patterns. For agentic systems the transaction is no longer the right primitive; Honeycomb shipped timeline, a trace-of-traces spanning a supervisor agent, its sub-agents, their API and storage calls, and a customer conversation that may run for hours. Agents querying observability data tend to bypass processed telemetry and go upstream for richer intact data, because the relationships are what matter and traditional telemetry just fills the context window with junk.
The second edition of Observability Engineering is a full rewrite — 600 pages vs 250, six parts, with a third of the book on observability governance for leaders: an open letter to CTOs, quantifying observability as investment vs cost centre, driving change without authority, build vs buy vs open source, and vendor partnerships as a durable senior-engineer skill.
Careers
Managers and directors: everyone gets to be hands-on again, and there is no excuse not to be. Middle management is essential — sense-making and context-giving — but if you're a director feeling the squeeze, going back to IC for a while is well respected. Anxiety and excitement are physiologically nearly identical; the difference is agency, so run towards the waves.
Get AI on your resume now. People with two or three years of AI engineering experience are in the best job market ever; people starting from zero after a layoff are struggling to get the benefit of the doubt.
Juniors will be fine if we hire them. The kids are cooking — they don't know what the SDLC is and are shipping anyway. Precedent: after the iPhone, the domain experts were 19-year-olds and the staff engineer was 22.
On AI fatigue: name which variety (slop in your inbox, hype, or "doom trolling" from lab CEOs — Cal Newport's term) and then take back control with local experiments — no AI-generated PR descriptions, no AI on Wednesdays. Don't wait for top-down permission; run it and report what worked.
Books she recommends
Catastrophe Ethics (Travis Rieder) — no framework saves you; educate yourself and decide what matters. More Everything Forever (Adam Becker) — a demolition of singularity/EA/accelerationist infinite-growth thinking, on the grounds that the one thing we know about exponentials is that they end.