Josiah Truasheim

Anthropic releasing code review agents that are billed at average of $15–25 Brilliant business model. 1. Make developers forget how to code by generating it all for them 2. Bill them per PR to verify the code nobody understands anymore A bit like an optician smashing your glasses then charging for the eye test This marks the beginning of the end of the subsidized inference era It will only go higher

18

1 Comment

Coding agents are amazing until they confidently say “done” after breaking half your app. Over several weeks in October last year, I explored Windsurf IDE as a coding agent in a spec-driven (spec-flow) workflow. I also compared multiple models (e.g., Anthropic Opus, OpenAI GPT-4, and others) using Spec Kit and spec-kit-mcp. I wrote down my thoughts after reflecting my experience on babysitting my coding agent and how I think this could be improved: - driver + reflective verifier agent - vertical slices over layers - event-driven slice boundaries - architect agent for contract changes - tight compile/run/test/fix loops I haven't really done any literature review, so the mentioned points may be well known. If you have any, I’d appreciate pointers. And I'm curious... - What failures are you seeing most? - What measurements did you implement prevent them? #AICoding #CodingAgents #DeveloperProductivity #SoftwareArchitecture

5

2 Comments

Have you thought about what Software 2.0 project repositories might look like? One path is where "our project" is just a set of .md files. Things like AGENTS.MD and companion files describing how code should be generated. You clone the repo, fire up Claude Code/Copilot/Codex, go for a lunch, come back and have a working application. The only problem I see with this... is repeatability. Every regeneration produces new code, different from the previous one, introducing new bugs. This problem seems solved by diffusion models in image generation — we have a seed, and if we use the same one, we get exactly the same image. Simple. There's an interesting paper on Hugging Face as "1# Paper of the day" titled "dLLM: Simple Diffusion Language Modeling". In short, we get a framework for converting existing models to a "diffusion" variant. So what? We get repeatability and predictability in code generation. That could be an interesting direction, right? 👇 https://lnkd.in/dTMRkdSW https://lnkd.in/dhTMJcDc #AI #Agentic #Future

1

1 Comment

🚀 LangChain / LangGraph Update – March 10, 2026 The latest release in the LangChain ecosystem introduces improvements that make building AI agents and workflow orchestration more reliable and developer-friendly. 🔗 Read the full changelog: https://lnkd.in/gghC6dSc #LangChain #LangGraph #GenerativeAI #GenAI #AIAgents #LLM #Python #AIEngineering #MachineLearning #AIDevelopment #RAG #AICommunity

17

This new TOON approach actually reminds me a bit of a CSV file — lean and purpose-built 😅 Here’s my take: It will naturally consume fewer tokens because it literally contains fewer tokens. The real question isn’t efficiency — it’s adoption. Let’s see how it holds up once the initial hype settles. JSON isn’t going anywhere. It’s an interoperability standard with decades of ecosystem maturity. TOON feels like a specialization — optimized for model I/O and token efficiency, not for general web communication. In a way, it’s the CSV of LLMs: stripped down, efficient, and designed for machines rather than humans. If adoption grows, it won’t replace JSON — it’ll coexist with it, just as YAML, Avro, and Protobuf each found their niche. Different tools for different layers of the stack.

3

2 Comments

Is JSON out and TOON in? (not a chance, but the difference is interesting) My feed has been flooding with these TOON vs JSON graphics and I can see why. Look at the token savings that is occurring with this new way of communicating data. one is built for a universal language for all HTTP requests, the other optimized for LLM token consumtion. Is one better than the other? I could also ask is a hammer better than a screw driver? the clear answer is no, because they each serve different purposes. But the question is, if TOON continues to be adopted more and more, would JSON become outdated and eventually deprecated? Let me know your thoughts! // p.s. TOON is kinda the better acronym ngl...

339

186 Comments

The project I mentioned in my last post works great, but it definitely took longer than I expected to get everything tuned. So for my next build — a new controller for a spin coater — I tried a different approach. If you haven’t seen one, a spin coater is basically a machine that spins a substrate at controlled speeds so you can lay down a thin, even film. Commercial ones are expensive, so I built my own with a drone motor, 3D‑printed parts, and an ESC. Up until now I was driving it with an RC transmitter, but I needed proper ramp‑up/ramp‑down curves and accurate RPM control. This time I used Google Gemini, and it took me about eight hours to go from idea to a fully working system: firmware, UI, PID tuning, ESC calibration, and support for pro‑level spin profiles. I shared both the prompt and the generated code in this repo. 👉 https://lnkd.in/gWu98ejp But the real story here isn’t the project — it’s the workflow that’s starting to emerge: My LLM‑Driven Engineering Workflow: • I start by having the LLM help define the module’s specs. It’s like a design review partner that never gets tired. • Once the specs look good, I ask the model to write the prompt that will generate the code. • Then I take that prompt to a second LLM and let it critique and improve it. (It usually says “this is already solid, but…” and then gives genuinely useful tweaks.) • I run the refined prompt in my code‑gen model and save that prompt as part of the project. • And yes — the model always misses something. That’s normal. I issue follow‑up prompts and keep them in a “prompt log.” • The magic happens later: when a new LLM version comes out, I can re‑run the same prompts and instantly get a better implementation without rewriting anything. And this is where things get interesting. LLMs are about to change code reuse in a big way I’m realizing that the old idea of “reusing code” is going to shift. Sometimes it’s not worth pulling in a third‑party library anymore. If the model already knows how to implement a PID controller, for example, it can generate a clean, purpose‑built version that fits your architecture better than a generic library. That’s a big deal. Because it means the thing we reuse isn’t the code — it’s the instructions that generate the code. We’re moving from: • reusing libraries → to reusing prompts • sharing code → to sharing the recipe that produces the code • maintaining a codebase → to regenerating it when the models improve The prompt becomes the durable artifact. The code becomes a snapshot — something you can always recreate, refine, or upgrade later. This feels like the beginning of a pretty big shift in how we build software.

6

1 Comment

Run Claude in "Dangerous Mode"... without the danger. I wanted to run some ideas using Raplh wiggum (Geoffrey Huntley ) plugin in claude code and leave it overnight without needing to baby sit terminal. Obvious answer is to run claude --dangerously-skip-permissions But I also wanted to run it safely in sandboxed directory The Solution: Linux Sandboxing (Bubblewrap) I (Gemini) wrote a wrapper script using bwrap to create a "Zero Trust" environment for the AI. 🔒 Layer 1: The entire OS and $HOME directory are mounted Read-Only. (The AI can read libraries but cannot delete your files). 🔓 Layer 2: Specific "holes" are punched in the shield for ~/.claude and ~/.claude.json so authentication states persist. ⚡ Layer 3: Only the Current Working Directory is mounted Read-Write. The Result: I can now run claude and let it act autonomously. If the agent tries to modify or delete anything outside the project folder, the kernel rejects the write request immediately. Try the script from here: https://lnkd.in/gjB93zJ7

21