Agentic Engineering Weekly for May 15-22, 2026

Jo Van Eyck

Jo Van Eyck

— 10 min read
Share
Agentic Engineering Weekly for May 15-22, 2026

OpenAI ran a negative 122% operating margin in Q1, Andrej Karpathy reframed his own term "vibe coding" as "raising the floor for everyone", putting it in sharp contrast with "agentic engineering". Giants like Google and Anthropic are hiring forward-deployed engineers by the dozen to embed at customer sites and make their own models actually do something useful. What we're left with under every theme this week is the same boring drum I've been beating: engineering discipline still matters.

My top 3 picks this week

Vibe coding got its tombstone, signed by Fowler and Karpathy

Martin Fowler added Vibe Coding to the Bliki this week with the maintainability caveat embedded directly in the definition: building software by prompting an LLM without looking at the code, best used for disposable software written for a limited audience. That last clause is doing a lot of the heavy lifting: it turns the definition into a warning label. Anthropic's latest employee Andrej Karpathy (who coined the term in February 2025) declared at AI Ascent 2026 that while vibe coding is about "raising the floor for everyone", putting it in sharp contrast with "agentic engineering".

What's actually being buried here is the assumption that LLM speed and prompts alone produce working systems. The escape clause Fowler wrote into the definition (limited audience, disposable software) is where the practice of vibe coding survives. Everything else gets the more demanding label of agentic engineering, which carries the obligations of reviewability, maintainability, security, and ownership. The same prompt that produces a working prototype in an hour produces a maintenance nightmare in six months. The split in terminology now is starting to reflect that.

Worth reading:

The Q1 2026 AI economics hit the books and the bills hit the inbox

This was the week the financial reality of the AI cycle stopped being theoretical. The Information reported OpenAI generated $5.7B in Q1 2026 with a negative 122% operating margin: losing $1.22 for every dollar of revenue, roughly $6.95B in the quarter. ChatGPT growth has stalled. Ed Zitron immediately documented Anthropic's claimed "first operating profit" as creative EBITDA accounting that excludes the actual cost of compute. These are the numbers the optimistic projections have been hiding behind for two years, and they are now public record.

Downstream, the bills arrived. The Register catalogued AWS and Google Cloud customers staring at five-figure surprise AI bills. GitHub Copilot raised AI token charges 10x to 100x in a single move. Theo showcased that the current "premium request model" is bleeding GitHub dry by letting him spend $40,000 worth of tokens for $40 of actual billing. One DIY Smart Code story tracked a single company's Claude bill going from $27K to $113K in one month. The OpenClaw creator is burning $1.3M monthly in tokens. The subsidies clock has audibly started ticking.

An entire escape-hatch subgenre emerged in response. Not Diamond shipped the first canonical "how to reduce Claude Code costs" guide. NetworkChuck switched off OpenClaw to a self-hosted Hermes setup on a VPS. Moe Lueker's three-tier model cascade runs at $8 per month. Anthropic split Claude plans in seperate buckets this week, resulting in higher costs for most customers still clinging to their subscription model. The implicit message in all of it: if your workflow assumed unlimited frontier-model access for a flat monthly fee, you're in for a rough summer.

Worth reading:

The "Forward Deployed Engineer" is back somehow?

Gergely Orosz's Pulse this week landed the framing: forward deployed engineering is heating up again, with massive demand for the role at Google, OpenAI, and Anthropic. The latest version of the FDE role is starting to look a lot like the classic consultant/solution-architect, only with agent fleets attached. Anthropic ran ten public FDE postings the same week. Mo Bitar's read on those postings was sharper: when the vendor whose pitch is "the model can do it" needs to hire humans to go embed at the customer's site and make it work, that is the most honest signal about where the limits actually are and why you'll be re-hiring your skilled engineers a couple of months from now.

The deeper reason this role is surging is that even with fleets of clankers and infinite token budgets at your disposal, you still need engineers who can hold a system in their head, translate business intent into technical scaffolding, and judge the agent's output against the actual problem. That is exactly the solution-architect-meets-staff-engineer profile. Sean Goedecke's How I use LLMs as a staff engineer in 2026 is the working practitioner's version: the LLM is an amplifier, not a replacement, and the skill is knowing where to point it.

Addy Osmani's Don't Outsource the Learning is the matching long-term argument and the reason this trend will keep accelerating. Trading future capability for present speed is exactly what happens when teams hand the keyboard to the agent and stop building mental models. The engineers who maintain their craft are the ones FDE roles are being designed around. If you are an experienced engineer wondering where to position yourself for the next two years, the deep-domain technical-customer-facing role is where the demand curve is actually pointing. Always has been, the ZIRP hiring spree of big tech just sweeped it under the rug for a little while.

Worth reading:

First AI-built zero-day is in the wild, and OWASP shipped the playbook

Google's Threat Intelligence Group published its 2026 adversarial-AI report this week documenting the first AI-generated zero-day exploit actually used in active attacks. An unnamed cybercrime group used an LLM to discover a vulnerability and build a Python script that bypasses two-factor authentication. This is the line we have been waiting to cross since the speculative research papers started appearing. It is no longer hypothetical, no longer "what if attackers used AI." It is in incident reports.

OWASP shipped the 2026 Top 10 for Agentic Applications the same week, peer-reviewed by 100+ industry experts. 404 Media obtained "Haotian AI," realtime deepfake software sold to scammers for face-swap during WhatsApp, Zoom, and Teams calls. The CEO-impersonation scam goes mass-market. The defensive side is racing too: IBM, OpenAI, Microsoft, and Anthropic all released competing AI-vulnerability-management tools (MDASH vs Mythos vs Daybreak), each one trying to detect and remediate automatically what attackers are now exploiting automatically. The good old security rat-race has just started to accelerate at the speed of tokens-per-second.

GitHub had its turn on Thursday: Theo's This is bad... covers the incident and his framing nails the cultural shift. Markdown files and on-device coding agents just became every hackers' new favorite attack vector. Maximilian Schwarzmüller titled his weekly update I'm getting tired, and that tone shift is the actual news: the community has stopped treating each incident as a unique event and started treating them as steady-state load. The OWASP Agentic Top 10 is no longer an interesting read. It is your minimum viable threat model.

Worth reading:

The productivity argument hits more real numbers — and they cut both ways

The 10x productivity argument has been running unchecked for two years. This week it ran into some more actual data. Abundly's Henrik Kniberg published concrete numbers. Gergely Orosz's Pragmatic Engineer Part 2 of the 2026 AI survey shipped: tradeoffs of AI tooling, why company-level adoption is hard, what changed in two years. Maddy Zhang's market analysis lands the headline contradiction: software engineering job postings are at a three-year high while companies continue to lay off. The labor market is sorting, not contracting.

The most concrete data point came from Barely Human Labs: two-thirds of companies that made AI-driven layoffs are rehiring, often within months. The developers they want most urgently are the ones who maintained their craft while the AI replacements were being tested, and those developers are not cheap. Block laid off nearly half its workforce citing AI investment; the rehiring already started. Andrew Murphy's AI didn't kill your junior pipeline. You did makes the matching long-term argument: companies stopped training juniors because someone saw an AI write a for-loop, and in five years they will be buying senior engineers from an emptied market.

The operational consequence shows up in Sam Newman's Dark Factories, Cobots, And The Potential Future Of The PR. AI-authored code means more PRs and bigger PRs. The review pipeline becomes the rate-limiting step. BCG Platinion's Dark Software Factory and Sander Hoogendoorn's From Excel to AI triangulate the same systems thesis: productivity gains are real, but the bottleneck moves to whoever has to make sense of the agent output. The honest read on 10x is that it is genuine and unevenly distributed, and the gain at the typing layer gets eaten back at the reviewing, debugging, and maintenance layers if you do not also re-engineer those.

Worth reading:

Code still has two audiences

Ian Bull's Sinks, Not Pipes: Software Architecture in the Age of AI makes a clean case: low coupling, high cohesion, and minimal side effects matter more now, not less, when agents are the ones navigating your codebase. Agents reason locally and miss architectural intent that lives outside the code. The discipline that used to be optional rigor is now a feedback channel for the agent itself. Bull's other piece (Why Embedding a JavaScript Runtime Inside an LLM Is a Big Deal) frames Anthropic's Bun acquisition as the unlock for accurate in-model computation, which is the supporting infrastructure story.

Mozaic Works's video commentary Reframing Code for the Age of LLM Assistants picks up Unmesh Joshi's What Is Code? from last week and lands the same point: code serves two audiences, the machine and the developer, and the machine-facing part is being commoditized. What remains is the developer-facing part, which is fundamentally a communication-to-future-readers problem.

The cumulative claim across these pieces is the durable one: architecture, judgment and taste are the surviving high-leverage skills in the agent era. Addy Osmani's Don't Outsource the Learning and Matt Pocock's 9 Ways AI Coding Has Rewired My Brain are the personal-practice angle on the same thread. The bug gets fixed by waving your magical prompt-wand; the mental model has to keep up. If you treat the agent as a decision maker, both the software design and your understanding decay. If you treat agents as a sharp tool that needs context, judgment and taste, both will survive and thrive.

Worth reading:

Quick Hits

Curated from articles, podcasts, and videos. Week of May 15-22, 2026.

Read more