Agentic Engineering Weekly for August 21-29, 2026

Jo Van Eyck

Jo Van Eyck

— 8 min read
Share
Agentic Engineering Weekly for August 21-29, 2026

Leave dozens of agents alone for long enough and they build themselves a legal system. Uber published the numbers on their software factory. Bill Gates is starting to get nervous. Legacy migrations are tractable today.

My top 3 picks this week

The harness is the thing

Models are commodifying. The skills, the extensions and the harness around them are where you can make the difference. I've come to think of this as building a habitat where humans and LLM tools can collaborate effectively. Same discipline as before AI, same questions about clear interfaces, but with a different tool that allows for insane speed. Kent Dodds' ADR video is a useful example I touched on as well: decisions die in chat, so write them down once and let the next agent read the answer instead of asking you again. I do this with architectural decisions in the large, but also with smaller feature-level decisions after planning/specc'ing/grilling. My ADR's have a long half-time, my specs tend to disappear after the factory has run its course and the team has absorbed the change in the system.

Swyx points out that models keep absorbing the harness into their weights. Whatever you build today gets eaten by models tomorrow. Addy Osmani's answer is to audit your agent files regularly and delete what the model now knows on its own, which is uncomfortable advice for any org who spent last quarter building context infrastructure and pausing shipping. My advice: build it anyway, keep it lightweight. Expect to throw it away with the next model generation, and design accordingly.

Worth reading:

Govern with fences instead of sandboxes

Ten weeks, dozens of frontier agents, one video game, and roughly $122k a month in tokens. Yegge's essay is another fever dream worth the read. Left to organise themselves, his agents didn't build an engineering process. They built a constitutional legal system. Rulings, case law, offices, enforcement. His argument is that this is what governing AI workers actually looks like, and that the sandbox model we've been reaching for is the wrong shape. Fences, not walls. Rules with consequences, not containment.

Johann Rehberger found a prompt injection against Claude Code's auto mode that he says works four times out of five. OCaml maintainers are now seeing exploit attempts within roughly ten minutes of a patch being posted for discussion. The gap between disclosure and weaponisation is shorter than a coffee break. All of this when Bill Gates picks this week to walk back his own "the risks are manageable" stance.

Worth reading:

Software factories at Uber and Ramp

More than 70% of pull requests at Uber now come from local or cloud agents, and lines of code per engineer has doubled year over year. Ramp went further and built its own coding agent, Inspect and is reportedly ahead of what the frontier labs ship. Twin Sun are running an actual dark factory in production, where agents write, review and ship with minimal human oversight. These are no longer mere demo's or thought experiments.

Read those numbers with your copy of Goldratt in hand. The theory of constraints says speeding up one station relocates the bottleneck rather than removing it. Uber doubling code per engineer tells you the coding station got faster. It tells you nothing about whether the line got faster. Find the bottleneck, exploit it, subordinate everything else to it, elevate it, then start again. Doubling output at a station that's not the constraint is how you build a pile of very expensive work in progress.

Addy Osmani wrote on whether or not you yourself will need a software factory. You get surprisingly far with a stock harness, several parallel sessions and specs with verification baked in. A factory earns its keep when you need consistency and a control plane. Zach Lloyd's pitch adds the part I agree with most: define the factory as code, version it, ship it with evals, and tune it on your own data rather than on this week's trending article.

Worth reading:

Migration is the first job agents finish end to end

Asana migrated off Enzyme in two weeks. Without AI that work would have been kicked down the road forever. Airbnb and Uber tell the same story. EVE Online just started moving 2.4 million lines from Stackless Python after sixteen years of not doing it. Everyone reported these as productivity wins. Why so many migration success stories? Because proving correctness is basically free.

The old behaviour is the specification. With these success stories, people object that they were cheating, that they had an oracle to compare against, so it was easy. But that's missing the point and laughing at your own stupidity. Building the verifiability layer, the golden master, is exactly where the work is if you want to succeed with these modernization challenges. Build a verification system, give proper direction, and agents rewrite it all while you sleep. Skip the verification system and you get plausible output nobody can check.

Which brings me to the sharpest warning of the week. Lada Kesseler trusts AI tests less than AI code. Her loop is deliberately dumb: give the agent one step, make it write the result to a file, read it back, go again. She calls it a centrifuge, spin fast enough and the stupid comes out. The failure mode she's guarding against is the one I expect to define the next year: Teams that don't have a real oracle and lack the expertise to build it out will generate themselves a fake one, call it a test suite, and ship against a mirage.

Worth reading:

Everyone adopted yet everyone is working harder

Linear's data says adoption is close to saturated and the hours didn't move. Meta explored cutting many teams by as much as 60% on the assumption that AI would absorb the work, then the CEO chickened out hours before the first layoff wave. Are we becoming more productive or are we just thrashing harder?

My own numbers land in the same ball park. Looking at a team I work closely with, pull requests per team member saw a 2-digit increase. What keeps me up at night is a pattern called cognitive surrender: colleagues outsourcing the thinking along with the typing. Taking pride in your own work is degrading into "someone else will look at it so I don't have to". That attitude won't age well. The people that rightfully worry about their professional career and the people adopting this mindset? A perfect circle.

Worth reading:

Expertise needs friction and we're removing it on purpose

Skill formation needs friction, and most people have spent the last two years removing friction as a design goal while building agentic coding harnasses. The Dreams of Code video takes the "writing code by hand is a waste of time" consensus seriously and then argues the opposite from where competence actually comes from. Mallika Rao, coming at it from the staff engineer angle, says the data structures and algorithms foundation everyone trained on is no longer sufficient on its own. Was it ever?

I ran a workshop on this with engineers ranging from zero to 10+ years of experience. We mapped four plausible futures. In all futures, every group independently concluded AI writes 100% of the code. What differed was the role left over for software engineers. The last generation of software engineer, the full-time reviewer, the factory builder. Learning on the job gets increasingly harder, while judgement and accountability stay human regardless.

When making gets cheap, judgment gets expensive, and taste is what separates the good from the generated. If you're working in unfamiliar territory, don't ask Codex to write the code. Ask it to teach you. Outsourcing thinking and skipping learning is a choice, not a consequence. If you don't feel like you have this choice today, it's time to find another employer.

Worth reading:

The model layer is commoditising

Multiple open-weight models shipped this week: Qwen3.8-Flash-Next, Ox Alpha aka GLM5.3 Flash. We are at a point in time where the floor has risen far enough, most agentic work no longer needs the frontier. Security-sensitive organisations can run their "Opus at home" now.

Anthropic's best model is losing ground to cheaper alternatives, even as annualised revenue climbs. Those two facts sit together fine, and Drew Breunig explains why better than anyone. Before Fable it felt silly to spend time improving your harness, because a cheaper better model would arrive and paper over your problems. Fable ended that. The cost was high enough that teams started thinking hard about which workload goes to which model. That's an interesting engineering problem if you're looking for one!

Worth reading:

Quick hits

Curated from 71 sources across articles, podcasts, and videos. Week of August 21-29, 2026.

Read more