
I read the source of Pi, Hermes and OpenCode to see how they route agent work. here's the routing ladder, what PostHog and Spotify learned, and prompts to audit your own agent. (I lead engineering at VerifyYou.)


Head of Engineering at VerifyYou. Empowering businesses through tech at Hickory Labs. Writing on engineering, AI & business.

I read the source of Pi, Hermes and OpenCode to see how they route agent work. here's the routing ladder, what PostHog and Spotify learned, and prompts to audit your own agent. (I lead engineering at VerifyYou.)


New models are incredible at zero to v0. The cracks show up around v0.3, once the project has real history.. old decisions, weird edge cases, code that exists for a reason nobody wrote down.

You can't prompt your way out of prompt injection. Any text an agent reads can try to give it orders. So the fix isn't a stricter system prompt. It's limits the model can't talk its way past: a proxy that blocks every site you didn't allow, files it can't see, no keys within reach. NVIDIA's OpenShell is built exactly this way. Worth a look: github.com/NVIDIA/OpenShell

Jesus Christ it’s Jason Bourne
Researchers have used reinforcement learning to teach a 5-tonne John Deere skid-steer loader to perform complex acrobatic manoeuvres autonomously. Trained in simulation and transferred to a real machine, the system can balance on two wheels, navigate rough terrain, spin, and control its bucket before landing safely. The research demonstrates the potential of AI-driven automation for heavy construction and industrial machinery. Video Credit: @leggedrobotics / @ETH Paper: wevlv.co/4kltYys #engineering #technology

I saw someone admit they tore down and rebuilt their software factory five times. That's five more paying customers than they have.

I wrote about what could replace it here 👇
I want my agent to shop for me. But how does the store know what I approved? And if I send 100 agents, how does it enforce “one per person”? Some thoughts on CAPTCHA and what comes next for trust on the agentic internet.


Proving you're human was the old question. Now agents shop, book and sign up for us. The real question is which human this agent is acting for, and whether they actually said yes. CAPTCHA can't answer that. Nothing on the web really does yet.

Dots on codex is functional but scuffed. Idk if anyone else experience this but I have an M5 and the app it’s so slow when opening the right side drawer for browser. Something is not right..

Switched to Bitwarden. Big regret. On iOS I have to unlock it like 3 times just to open my vault, and half the time it won't autofill my password until the 2nd or 3rd try after Face ID. What is going on 😩

Memory is honestly the most annoying part of building agents. Memories go stale constantly and managing them is a headache. This is a really interesting take: let the agent rewrite its own rules for what to remember and what to recall, then keep only the versions that actually make the output better. Give it a read 👇 x.com/jaseweston/status/2122…
Introducing: Self-Improving Memory 🧠✨📈 We let the model discover its own memory skills, then train them into its weights: - Meta-optimizes memory from the model's own runs with large gains. - Distilling these gains works too. Blog post: 📝 facebookresearch.github.io/R… 🧵1/6

Follow-up to my verification loop article: a repo you can try. Prompts to audit your verification loop, a small CLI that ties test results to the commit they ran on, and runnable CI, pre-push and Playwright examples. Adapt it to your stack. github.com/jschiezty/verific…
I pulled together lessons from the Claude Code team and our workflows at VerifyYou: from pre-push checks and CI to preview environments, agent review, and production monitoring. I included prompts you can use to assess your own verification loop too


I pulled together lessons from the Claude Code team and our workflows at VerifyYou: from pre-push checks and CI to preview environments, agent review, and production monitoring. I included prompts you can use to assess your own verification loop too


Life’s crazy… I’m watching AI grow up so fast in front of my eyes. Have we surpassed adolescence and in to teenage or are we already in adulthood?

I’ve been digging into evals while thinking about how to build a better agent harness. One thing I keep coming back to is that a successful outcome isn’t enough. We also need to understand what the agent did to get there, what changed, and whether we can trust the checks judging it.
Just published my top eval papers of 2026. The list covers different aspects of evals to focus on: durability, quality, robustness, efficiency, etc. It also covers new topics like verifiers, agent-as-a-judge, jev-as-a-judge, audit self-improvement, and more.

The idea of an AI workspace as a lab notebook really resonates with me. I don’t just want the answer or the code at the end. I want to understand the experiments, tradeoffs, and decisions that got us there. Looking forward to trying SolveIt.
We built a tool for our own internal team at AnswerAI to let us learn, explore, & develop mastery with the help of AI: SolveIt. Today, we’re releasing it to the public for the 1st time. fast.ai/posts/2026-11-16-sol…

Still figuring out what this means for our industry. But personally, I feel supercharged. More time with users, closer to the business, and a much shorter path from “this is a problem” to trying a solution. That’s the part I’m excited about.
A brief history of AI Coding since Nov 2025: Nov 2025: Opus 4.5 releases, OpenClaw gets built Dec 2025: community agrees that models get good enough to handle 95% of coding tasks Jan 2026: Ralph loops get popular, early efforts on loops/orchestration Feb: Everyone feels knackered, too much strategic decision-making, fears around cognitive debt begin Mar: Skills explosion as community converges around easy-to-share processes (superpowers already ahead of the game here) Apr: most SOTA devs haven't written a line of code in months May: "The unreasonable effectiveness of HTML" - more interest in richer review surfaces Jun: "Loop Engineering" wins vaguepost of the year, signals more hype around process Jul: Every talk in every conference becomes "here is my software factory" Aug: Opus 5 speaks incomprehensibly, interest in ASD-STE100 peaks Sep: Grok Bot gets popular, more interest in long-running "manager" threads which retain context and delegate work Oct: The promised speed gains haven't materialised, "dark factory" feels a long way off still Nov: No-slop manifesto releases, reading the code still appears essential TL;DR: no-one is writing code anymore, and we are still figuring out what that means