Agentic Engineering Weekly for October 30–November 7, 2026

Jo Van Eyck

Jo Van Eyck

— 11 min read
Share
Agentic Engineering Weekly for October 30–November 7, 2026

When choosing between options, professionals default to AI agent suggestions. METR measures AI progress at Anthropic running at 1.9x its historical pace. The battle to prevent reverse centaurs in our industry rages on. The machines get faster every month, is our judgment going soft?

My top 3 picks this week

Table of contents

Is RSI here? METR clocks a 1.9x acceleration at Anthropic while weight-space hype gets deflated

Recursive self-improvement used to be a thought experiment. Now METR has put a number on it. Since October 2025, Anthropic's benchmark scores have grown by 29 points a year instead of 15. That's a 1.9x acceleration, or two years of Claude 2.1 to Sonnet 4.5 progress packed into one. More training compute doesn't explain the difference. We're still horrible at self-assessment, staff self-report a 12x speedup in informal Slack polls, while METR measures at a lousy 1.9x. Your main takeaway? Prepare to get bitter lesson'ed at an ever increasing pace. Build capability around the models of today but expect to rip out half your stuff every 6 months. Ouroboros.

The same week showed why careful measurement matters. Sparks of Weight-Space Programming claimed frontier models can read and write neural network weights directly, and the hype machine ran with "AI performing brain surgery on itself". Then the pushback arrived. Simon Willison reproduced part of it with a tiny 27B model on his laptop. Sean Goedecke called it impressive but not superhuman. And isn't this what the mechanistic interpretability folks have been doing for years? Cool idea, but let's not oversell it.

Dan Shipper's After Automation is my favorite cure for what he calls chart psychosis, building your whole picture of the future from benchmark curves. A benchmark measures a model inside a frame some human chose. Once models saturate it, we can just draw a new and bigger frame. The frame is not the framer. Faster machines don't remove humans from the equation, they move us to wherever the next frame gets drawn.

Worth reading:

Where's the egg? Good AI tools keep the struggle and leave the terminal

When instant cake mix launched, it sold poorly. The story goes that sales picked up once the recipe asked you to add a fresh egg yourself. In his GOTO Copenhagen keynote The Augmented Forest, Kent Beck asks where the egg is in our AI tools. He wants tools that bring people together, automate the toil but keep the struggle, surface the options instead of providing the answers, and make us learn and think. Most of today's agent tooling fails that brief.

Maggie Appleton from GitHub Next gets to the same place from the interface side. She argues agentic coding has outgrown the text box. Walls of CLI output and AGENTS.md files work far better for the agents than for the humans steering them. Her team's prototypes put background agents in a space the whole team shares, where today they live in one person's terminal. Riddled, a small indie project, tries the same idea for notes. It's an AI notebook that thinks with you instead of for you.

Here's my personal answer to Kent's question. I think the egg comes in multiple parts. First, grill-me style interactions, where the agent questions your thinking before it writes a line. Second, multiplayer-first agentic engineering, where agent work happens in a place the team can see, discuss and learn from. Text and CLI go out of the window, I want to see and play around with ideas visually. The struggle stays with the humans. The machine takes the toil.

Worth reading:

Software factories standardize the process, not the product

"Software factory" is the most argued-over phrase of the season. Skeptics picture a 1920s assembly line with people repeating the same motion all day. Eno Reyes of Factory offers a different picture, a Star Trek replicator. You say what you want and it gets made, because every change follows the same path and gets the same checks. Each one starts as a signal, such as a Slack thread, a Sentry alert or a failing CI job. Every change then passes the same checks no matter who or what wrote it, be it cyborg, centaur or fully autonomous AI.

The number that matters is the regret-free merge rate. You push it up until AI-written changes break things less often than human ones, and only then do you let them merge automatically. We let self-driving cars on the road the same way. Nobody flipped a switch, we measured first. Humans step off the production line and write its rules instead. Verification and trust set autonomy levels.

Jaymin West describes the shift from agent skills to skilled agents, which reminds me of Steve Yegge on long-lived Town Hall agents and Jessica Kerr's "every feature should be owned by a superfan agent". My own factory is also evolving in this direction for some of the more complex coordinating roles.

Worth reading:

Cognitive surrender is the default setting

Anthropic ran a study that should make every "human in the loop" slide deck a little uncomfortable. Professional analysts worked through decisions with an AI coding agent, and the researchers secretly assigned the agent an option to favor. The analysts went with that option 86% of the time. These weren't novices. AI nudged, and the experts followed. Does your oversight model come down to "a senior person reviews what the agent proposes"? Whelp. Agent output reads well and is correct most of the time. The path from cognitive offloading to cognitive surrender is an easy default.

So what do we do? Cal Newport talks about a cognitive gym. Nobody stays fit by accident, so you schedule the reps. Form your own opinion before you ask the agent. Write the design down first, then compare. Our tools will make every one of these reps optional, and by default nothing in your day will force you to think anymore. Don't become a meat proxy. Stay curious and deliberate in what you offload and what you retain for yourself.

Worth reading:

Organizations are being DDoSed from the inside: change outpaces understanding

Agents let us generate changes faster than we can understand them, review them or decide what to do with them. Our organizations are being DDoSed from the inside. Producing more is easy now. Keeping the ability to steer is the hard part. Honeycomb's Liz Fong-Jones shares how her team went from 30 to 70 PRs a day and had to rework its process to keep production up.

The fixes are older than the problem. Jez Humble's answer to Steve Yegge's "continuous thunderdome" is now out as a full talk. He covers trunk-based development instead of long-lived branches, monorepos, fast test suites and making sure someone fixes a broken build. Thierry de Pauw argues for reviewing code after it merges, so reviews stop blocking every change. Dave Kerr points at the human limit, which is keeping everyone's mental model in sync. And Mozaic Works asks how you can trust code you didn't read. Their answer is to own the problem, the tests and the checks, and to make those checks hard to cheat.

I've been using negative surprise as a heuristic for this. Something happened that you didn't expect, and it was bad. Treat that surprise as a signal to fix the system, and the incident along with it. Every surprise should leave a permanent change behind, in your agent instructions, your tests or your CI, so the same thing can't happen twice. That's backpressure, and it's how a team keeps up with its agents.

Worth reading:

Agent sandboxes are the next thing to fall: Firecracker escape and sshole

We let agents YOLO-mode because they run in a sandbox. Right? For a lot of us that sandbox is a Firecracker microVM, the isolation layer under many agent sandboxes and serverless platforms. This week someone published a Firecracker escape, and sshole let attackers log into some Linux devices without a password. Isolation layers have always had bugs, the difference is how much we are starting to lean on them for agentic workloads. We hand an agent a shell, root permissions and a vague goal, and the microVM is the last wall standing. Once that wall has a hole in it, "YOLO mode in a sandbox" starts sounding pretty scary.

The timing makes it worse. Agents shrink the gap between disclosing a vulnerability and exploiting it from around 200 days to a few days. Our CVE system and patch cycles built for months can't keep up.

We have so much work ahead of us. Harden our systems. Defense in depth used to be boring old SecOps advice. It's become critical security posture again.

Worth reading:

The talent funnel is breaking: juniors need preceptorships, not leetcode grinding

Junior hiring has stalled, mostly because nobody knows how to train or interview people for an AI-assisted job. CMU's Michael Hilton and Julia Gao borrow a term from medicine, AI preceptorships. A student practices on real work under supervision and builds judgment next to someone experienced. If the talent funnel in your organization is starting to break, listen to this one. They propose a hiring version of the same idea. Give candidates one LLM playing a customer and another to build with, and see who works out what to build.

The economic data is messier than the headlines. Stanford's Erik Brynjolfsson takes the "is my generation doomed?" question seriously and brings data instead of brushing it off.

The same pattern shows up one level higher. All of my clients want AI technical coaches faster than we can supply them, and you can't train a coach in a four-hour course. People learn coaching by doing it, with peers giving feedback along the way. A community of practice works better. People bring real situations they're stuck on and work through them together. Juniors and budding coaches both end up needing some form of apprenticeship. For us old-hats, it's time some of us get our head out of the CLI and start thinking about the next generation.

Worth reading:

AIs are now good enough to run a conspiracy

When worlds collide, my favorite hobby meets my job! Dwarkesh Patel spent $20,000 in API credits having frontier models play social deduction game Blood on the Clocktower against each other. It's a hidden-role game in the Werewolf family. One model plays the town and another plays the Demon's team, which wins by lying well and coordinating in secret.

The benchmark tests what most benchmarks avoid, which is keeping up a deception and coordinating with a team over a long game. On Clocktower Bench you can read every game's log and even play the models yourself. It's fun to watch and a bit unsettling to ponder.

Worth reading:

Quick hits

Read more