• Tencent Hy4 preview released
  Don’t get confused by the title, “preview” is part of the model name

• Do you need a software factory?
  There’s a lot to think about regarding project and developer needs and this article explores some considerations

• The Harness Is the Thing
  A proper example of customizing your harness

• What is reasoning
  A concise but good explanation of the concept

• What is a harness
  Trying to be precise about what the harness is and what falls into its responsibility

• Fast and Hard Code
  Another interesting deliberation on how coding is changing

• Why your local LLM feels dumber than it is
  A very cool exploration into local inference

• Quick impressions: A week of using Codex more than Claude
  It’s always hard to do these comparisons but it’s good to get a sense check of how other people perceive the different agents.

• Pi’s Minimalism Is Its Advantage
  The models have gotten smarter. Perhaps bloated harnesses are not the way to go.

• Introducing Shieldstral.
  A model targeting content moderation

• Introducing Muse Code and Muse Spark 1.2
  Meta brings their own coding agent and a new model to the race

Videos:

• jellyware and more thoughts on jellyware
  The amp team had a get together and these are some thoughts coming out of one of the team members on where software is going. Obvious grain of salt recommended.

• In defense of not understanding your codebase
  The article explains why partial understanding of a codebase might not be the end of the world and touches on the differences between “pure“ and “impure” software engineering

• We rewrote our agent to run entirely in a Durable Object with Pi, Agents SDK and Code Mode
  Cloudflare was able to move their agent away from Linux VM sandboxes to Durable Objects with R2 and reduced latency and costs

• Our position on open-weights models
  Dario Amodei (Anthropic CEO) explains Antrophic’s stance on open weight models and argues for collaborative safety and alignment testing regardless of licensing

• Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
  A very interesting writeup about how exactly HuggingFace got breached by OpenAI

Videos:

• /wayfinder: Nothing is too big to plan anymore
  Another iteration on the already great skills collection from Matt Pocock. The new wayfinder skill helps you plan complex roadmaps that might exceed the limit of a single context session

• Codeberg is taking a defensive stance against LLMs
  There’s projects that are against AI generated code, but it’s not common (at least yet) for a git provider to go in that direction.

• Claude Opus 5 released
  Half the price of Fable for close to the same performance is the big claim.

• The new rules of context engineering for Claude 5 generation models
  The Claude team shares their lessons in regards to context

• Nothing Works and Everyone Is Euphoric
  A good depiction of a feeling a lot of us are having

Videos:

• All developers should read less code
  The best idea might be for you to ignore the title and watch the video. It makes good arguments for why you should use AI to generate more code and not all of it is worth the same attention, at least as far as reviewing it goes.

• New Grok model released but it uploads your whole repo
  You might want to try it out, but not on a project with secrets in ‘.env’ files.

• Introducing Inkling
  The company founded by the ex-CTO of OpenAI Mira Murati has released their first model

• Kimi-K3 released
  A new model promising to compete with state of the art models at a cheaper price

• AI usage in Linux kernel development
  Linus takes a stance in favor of AI usage on the Linux mailing list

• Rewriting Bun in Rust
  For anyone who is curious about how the rewrite of Bun from Zig to Rust was done in detail. The use of an adversarial review agent sounds promising.

• Anthropic's Method to Losing Goodwill in a Few Easy Steps
  A nice breakdown of all the shenanigans that Anthropic has pulled over the last few months. Our favorite quote: “Dario and Boris have us convinced that ‘coding is solved’ with their loops. But microwaves didn’t solve cooking.”

• Continual Learning for Agents
  Replit presents their framework for continuous learning of their customer facing vibe coding agent.

• GPT 5.6
  Another week, another model. At least OpenAI tries to bring more structure into their naming scheme. Sol (Sun) → Terra (Earth) → Luna (Moon) in descending order of capability.

Videos:

• I Built an LLM From Scratch
  CJ (Co-Host of Syntax.fm) is going into detail on how he built his own LLM. For anyone who wants a deep dive on how the tools work that we use everyday.

Before we get started this week we want to clarify something:
Our posts are written without any use of AI.
Nowadays we get bombarded with AI generated content all around and believe in the principle that Armin Ronacher introduced: “Content by humans, code by clankers”

• Qwen 3.6 27B is the sweet spot for local development
  Qwen 3.6 is taking the local coding agent community by a storm. Users are claiming that this is the inflection point for local models that we experienced last year with frontier models like Opus 4.5 last year.

• Claude Sonnet 5
  Anthropic released a new major version of their Sonnet class models. Performance is claimed to be on par with Opus 4.8 in certain tasks for 1.3-1.5x the cost of Sonnet 4.6 thanks to a new tokenizer.

• Agents in Orbs
  If you listen to the developments in Silicon Valley, cloud agents are becoming all the rage. This seems like a very nice DX for this workflow.

• Putting an Agent in an Orb
  A follow-up to the link above. If you wondered how the sausage is made, this article is for you.

• Better models: worse tools
  The newest Anthropic models are causing faulty tool calls in pi. The cause: heavy reinforcement learning bias towards a closed source harness (Claude Code).

A lot has happened this week again and we don’t think we are going to run out of content any time soon. We appreciate the support, engagement and feedback that we received ever since starting this journey a month ago.

• The Coming Loop
  A reflection over loops and expertise with them.

• Will It Mythos?
  While it’s hard to know how much about Mythos is pure marketing, this is a great effort in that direction.

• Introducing OCR 4
  New Mistral model focused on extracting content from documents. And they claim it’s compact enough to deploy on a single container.

• US government approves GPT-5.6 customer by customer (full post behind paywall)
  Politics involvement in AI, at least in the US, will continue for the time being.

• OpenAI CEO claims 750 token/sec coming soon
  Crazy performance improvement.

• OpenRouter announces Fusion API
  Turns out you can combine models to try to have better results.

• SpaceX acquires Cursor for 60 billion USD
  The company that a lot of people dismissed as being “just a VS Code fork” clearly has way more to it than that.

• Vercel announces eve
  An easier way to make agents. Pick an LLM and start configuring around it.

• Flue framework 1.0
  Similar to eve in goals, but more agnostic. Brought to us by the same people that made Astro.

• We don’t write code anymore
  A field report over how a team’s way of working has evolved

• WTF Is a Loop? Peter Steinberger vs. Boris Cherny
  Quick “what’s a loop?” explainer + some Steinberger vs Cherny internet debate energy.

• Loop Engineering
  Loop engineering = set up an agent to iterate on goals so you’re not babysitting prompts all day.

• Lines of Code Got a Better Publicist
  “AI wrote X% of our code” is the new lines-of-code flex.

• Statement on the US government directive to suspend access to Fable 5 and Mythos 5
  The empire strikes back and the US government mandated Anthropic to take down their new flagship models until further notice.

Videos:

• Learn anything with the /teach skill
  Great demo of the /teach skill, basically “learn X” and the AI walks you through it.

• How LLMs actually work

  A nice walkthrough of how modern transformer-based LLMs work. Advanced math degree not required to follow along.

• A harness for every task: dynamic workflows in Claude Code
  The announcement, why and how of Claude Code’s new “dynamic workflows” feature.

• The Orchestration Tax
  The costs of running many AI agents when your own cognitive bandwidth (judgment, review, merging) remains serial and limited.

Videos:

• PI Architecture EXPLAINED | Agent Loop, Tools, TUI and More
  Deeper dive into how the pi coding agent works under the hood. The author of pi gave a compliment on the accuracy of the content.

• Testing How Much AI Has Improved In 6 Years
  A fun overview putting into perspective the evolutions up to this point, with some cool practical examples.