Agentic Engineering Weekly for October 22–30, 2026

Jo Van Eyck

Jo Van Eyck

— 10 min read
Share
Agentic Engineering Weekly for October 22–30, 2026

Anthropic deleted the coding interview, Amazon rewrote what it expects from a developer. Two AI labs claimed to have broken a 27-year-old pillar of complexity theory. The job, the hiring bar and the limits of computer science all got renegotiated in a single week. Let's continue to make sense of it together.

My top 3 picks this week

Agency is the new seniority: Anthropic and Amazon rewrite the hiring bar

The coding interview no longer provides signal. The coding round is gone from Anthropic's loop altogether. Amazon published a revised role guideline for its software development engineers: what matters is the outcome, not which tools you used to get there. Two very different employers, one direction. If the model can solve any whiteboard puzzle in seconds, the whiteboard puzzle measures nothing. So what do we hire for now?

One answer is hunger. Curiosity, drive and the willingness to try ambitious things now beat years in the industry. I find this a surprisingly optimistic take for junior talent. Sean Goedecke makes the same point: most of us are bottlenecked on noticing what an agent could do for us, not on what it can do. He lets his agent fix his broken dev server. Small thing, but how many of us still debug that by hand out of habit? Juniors do not need to unlearn 15 years of habits. In a follow-up Sean pushes back on the "there's nothing to learn, just tell it what you want" crowd: getting "good at AI" takes deliberate practice, like any other craft.

My favourite line of the week comes from the Amazon guideline: "If your work does not benefit from current AI tooling, that is a finding, not a failing". The judgement Amazon asks for cuts both ways: use the tools ambitiously, and explain clearly where they don't help yet.

Worth reading:

Eulogy for the software engineer

If you watch one thing this week, make it Dave Kiss's Eulogy for the Software Engineer. He holds a literal funeral on stage and buries three props: headphones, a notebook, a keyboard. Then he shows what's still in our hands. Going all the way in. Writing down what "done" looks like. Standing behind what ships. It's a work of art, beautifully delivered. It puts into words the complex mix of feelings so many of us have carried for the last two years, and it ends with a clear way forward for individuals young and old, and for the profession as a whole.

The grief is everywhere right now. Glyph argues that programmers are artists and should resist AI the way artists do. Syntax gets mail from a listener who thinks the show has gone too pro-AI. Soft Skills Engineering fields a question from a lead asking how to keep their team from "losing their soul to AI." The part agents can't take over is finding out what people actually need.

Aaron Francis has a sharper name for what's left: eating the pain. They point out that pain hasn't disappeared. Someone eats it: you or your customer. Good engineers, teams and products absorb the pain in place of the customer and pour care into the product. That's how you stand out in a sea of vibe-coded software. The floor just rose for everyone, the only place left to stand out is at the ceiling. Increase your vertical ambition: don't build more features, eat more of your customer's pain. Depth compounds into breadth.

Adam Mastroianni's Claude, don't touch my lump adds another perspective. We're not wrong to worry. But we are wrong to think there will be an end to worrying. The anxiety doesn't resolve once we've "adapted", because the ground keeps moving from under us. Models keep eating our carefully crafted skills and harnesses. That's a strangely freeing thought: stop waiting for the feeling to pass before you act. Bury the props, keep the practice, and start moving with the flow.

Worth reading:

AI is proving things no one understands: the Exponential Time Hypothesis falls

On October 16, both OpenAI and Anthropic claimed their models had broken the Exponential Time Hypothesis, a 27-year-old foundation of complexity theory that a lot of hardness results quietly lean on. OpenAI's write-up lists results on the Kakeya conjecture and quantum chromodynamics alongside it. No, this doesn't quite mean P = NP just yet. But ETH is a big deal: plenty of "this problem can't be solved faster than X" proofs assume it. Pull it out and you have to re-examine everything we've stacked on top.

Both Hank Green and a three-hour Lex Fridman episode with mathematicians all circle the same question: what is a proof worth if it's verified but no human can explain why it's true? If no human can build and generalize from the results? Maybe Terence Tao is right when he says this era is teaching us our idea of intelligence was never accurate in the first place.

Software engineers should recognise this pattern. It's similar to comprehension debt, the gap between software that works and software we understand, now at the scale of an entire scientific field. Mathematicians are about to run the experiment we're all running in our codebases: can we build on top of something nobody understands? I'm on the edge of my seat watching how they handle it. Their answers may well become ours, and vice versa.

Worth reading:

Trust, not code, is the binding constraint

Kent Beck put it in one sentence this week: teams accumulate code faster than they accumulate trust. GitHub's Octoverse numbers show the scale: merged pull requests tripled in under a year. The code is there. The confidence to ship it isn't.

So how do we grow trust at agentic speed? The Claude Code team shared a concrete answer. Once coverage and automated deep reviews were in place, they started optimising for the minimum amount of human review time that doesn't raise reverts or incidents. That turns "how much review is enough?" from a feeling into a measurable trade-off. Honeycomb goes a step further: spend the AI dividend on rigour you could never afford before. Property-based tests, formal verification and strict architecture checks all suddenly become cheap.

None of this is new. Safety-critical teams at NASA, in automotive and in avionics solved "code faster than review" decades ago: classify risk by consequence, write requirements as contracts, prove them with evidence. Not every module deserves the same scrutiny. A typo in a marketing page and a bug in a payment flow don't have to be subject to the same level of scrutiny. Matt Pocock's version is the most down to earth: get from one agent PR to a factory one trust increment at a time.

Worth reading:

Wheelhouse vs Counting House

"Software factory" is the phrase of the season, and it's already splintering. Steve's own Wheelhouse is the prototype developer-side factory: agents with memory and broad access, building software. His Counting House is its mirror image, the pattern he sees emerging in enterprises that put AI on millions of customer contacts a year. Stateless clerks. No credentials. No memory. Every write checked by an orchestrator. Every model promoted intent by intent on the strength of its evals, like a new hire getting more responsibility as they bo about proving themselves.

That last detail is the one to steal. Most teams roll out agents all-or-nothing. The Counting House earns autonomy one narrow task at a time, with evidence. It's the same trust-building idea from before, applied to operations instead of code.

Not everyone buys the metaphor. Joe Maddalone argues factory thinking drags outdated mental models along with it. Patrick Debois chases a dream factory: not just faster cycle times, but dreaming up ideas we could never otherwise do. Amp makes the case that culture, not process, is what stops slop and unlocks real progress.

Which should your organisation start with? Steve argues that while everyone is busy chasing the factory, most orgs should start with a Counting House instead.

Worth reading:

Text in the CLI is not the UI endgame

The terminal continues to be -quite baffling to me- the default home of agentic engineering. It's a terrible place to live and work. A 40-step plan scrolls past, you type "looks good, but change step 12 and drop 17", and hope the agent maps your prose back onto the right lines. Feedback detached from the thing it's about is lossy, walls of text are my personal version of hell. Yesterday a colleague demoed plannotator this week. Once the agent finishes planning, a visual browser view opens where you delete, insert, replace or comment on individual steps, then approve or send structured feedback straight back.

It's not a one-off, it's an emerging pattern. Bouke's remark does the same for markdown: threaded comments, resolution status and read-marks live inside the .md file, anchored to the paragraph they're about. You comment, the agent watches the file and replies in place. No chat window, just right-click and fire off async work. Kun Chen's Lavish embodies that exact same idea, with the tagline "HTML is the new markdown." Click an element in an agent-generated page, annotate it, sketch on a diagram, and the agent polls for your feedback. No more screenshots plus a paragraph of explanation.

Notice the shared pattern: the visual artifact becomes the conversation, and a tiny CLI loop (monitor, poll, approve) carries feedback back to your text-hungry agents. My bet: the chat transcript turns out to be the scaffolding of agentic UI rather than the UI itself. Try one of these on your next grilling session and count the round trips you save. You're welcome.

Worth reading:

  • plannotator: Annotate an agent's plan step by step before it writes a single line; works with most major coding agents. (article)
  • remark: Comment threads that live in the markdown file itself, so humans and agents share one source of truth. (article)
  • Lavish: Click-to-annotate review for the HTML artifacts agents increasingly produce. (article)

After Black October: attackers move at machine speed, security teams at meeting speed

Last week's breach wave is turning into a rethink. Gergely Orosz counts a month's worth of security breaches landing in a single day and asks whether things will get worse before they improve. Caleb Sima, former CSO at Robinhood, has seen an AI attacker up close: it moves at machine speed while most security teams still move at meeting speed. His conclusion is uncomfortable but hard to argue with: Only automated defence can keep the pace.

The labs have noticed. Google led its Gemini 4 Carbon launch with state-of-the-art cybersecurity capabilities, ahead of everything else. When the headline feature of a frontier model is security, you know which way the frontier arms race is heading. If your patching process still involves a change advisory board meeting, now's a good time to ask what it would take to fully automate it. We're living in the science-fiction movie folks.

Worth reading:

Quick Hits

Read more