Andy Nogueira

Building AI-native software, early days, heads down, deep focus · prev. founding member @informalinc · prev. @Autodesk Research

[Andy Nogueira @andynog]

@buchmanster @informalinc @buchmanster Godspeed! The future belongs to those who transform dreams and ideas into reality, reducing pain points and friction. All the best to the @cyclesmoney team

[Informal Systems @informalinc]

We're very proud to see the Cycles team's latest funding round. We're now all in on growing our spin-outs, and excited to continue supporting the security of the Cycles tech stack.

[Andy Nogueira @andynog]

Using a trackpad on a Mac with a large curved monitor is an absolute game-changer for juggling 4 or 6 simultaneous agentic sessions with ease!
🖥️↔️🖥️🤖🦾👨‍💻

[Andy Nogueira @andynog]

This article is insightful. It brought to mind a podcast episode featuring Professor Dan Boneh, who discusses this topic with authority. A notable point is the comparison between the AI top-down approach and Blockchain's bottom-up approach to addressing the human aspect, with a focus on proofs. Midway through the episode, he delves deeper into Worldcoin and C2PA, technologies present in some Sony cameras. Although the episode is dated (2023) in AI terms, Boneh effectively captures the current landscape and future directions.

https://t.co/See2uTMuxc

[Andy Nogueira @andynog]

I am a long-time Claude user (>200 million tokens), and my AI-native solution has been built entirely using Claude models since Opus 4.5.

Recently, I’ve noticed some trends of people switching models and harnesses, but honestly, I’ve always believed that both models and harnesses will improve over time across the board. What really makes a difference in my experience is how you adapt your customization (skills, hooks, agents) as the models and harnesses evolve.

As some others have pointed out, my main focus is also on maximizing token efficiency. I’ve always been a Pro Max user.

I have a custom status line that tracks effort, model, tokens used, caching, and 5-hour and 7-day limits, including when the limit is during peak hours. I also monitor daily token consumption to manage the rate effectively. My goal is to distribute tokens throughout the day to prevent hitting limits or burning out. I believe this session tracking is key to understanding token usage.

One thing I consistently do is maintain my own customization. Files like skills and CLAUDE.MD tend to grow over time after updates, and without pruning, they can become unwieldy. I also have hooks to detect file bloat and prune skills, making sure each session checks tracking files or changelogs. This way, I stay aware of drift or bloat. When I address these issues, my token consumption returns to a healthy average. I don't use MCP, and most of my setup is in Markdown files.

Additionally, every time I upgrade Claude, I ask myself, “What’s new in Claude [version],” search online, and see what can be leveraged. Often, there are improvements that benefit my custom skills and hooks, and I implement those.

Basically, I wonder whether many people are simply switching models because they’re not properly maintaining their harnesses and are unnecessarily over-consuming tokens. I'm not sure if switching is always the best approach. From my long-term experience, building on a consistent harness and continuously customizing and evolving it makes Claude more powerful over time, even if the model doesn't change. Sometimes, it starts to drift, but due to how I've built my customization, I’m always pushing back, detecting gaps, and using that feedback to improve future workflows. That approach seems effective.

So, I believe that educating people about this and helping them understand how to maintain their customizations (or offering tips) could improve token efficiency and, ultimately, customer satisfaction. Still, offering higher token limits is very helpful. This promotion, running until mid-July, has shown good results, and I hope it can be extended. 👨‍💻🤝🤖

[Andy Nogueira @andynog]

@paulg Scarcity that is not recognized collectively may go unnoticed economically.

[Andy Nogueira @andynog]

I really appreciate the nostalgic charm of the green screen interface on my agent's that resembles dumb terminal sessions. I know many enjoy the Claude desktop app, but I find comfort in the old-school feel that reminds me of earlier times.

[Andy Nogueira @andynog]

@ClaudeDevs, it would be more effective to increase the weekly limits rather than simply resetting them. For instance, my next weekly reset is in less than 24 hours, so this reset doesn’t do much good because I won’t be able to use the tokens. It would be much better if this extra capacity were added to my upcoming weekly allowance so I can spread it out more evenly.

[Andy Nogueira @andynog]

The more I observe my agents delve deeper, the more captivated I become. Witnessing their progress and knowing our efforts are clearly paying off is an exhilarating and rewarding experience.

[Andy Nogueira @andynog]

Yes, depth truly transforms the game. From my experience, it's what propels you forward more than spreading yourself thin. When you build with focused discipline and genuine depth, you create a strong moat.

[Aditya Agarwal @adityaag]

https://t.co/AoyPtuI9jx

[Andy Nogueira @andynog]

This is arguably one of the most compelling and precise articles I've come across lately, brilliantly elucidating how AI-native companies function and the exciting direction they're heading. It resonates deeply with my own developments and reinforces my core thesis.

[Luke Sophinos @lukesophinos]

https://t.co/9JEEPjtdnw

[Andy Nogueira @andynog]

Having operating constraints forces you to adapt and continually refine your workflow through feedback. Without constraints and having unlimited resources, you miss opportunities for efficiency, which is crucial for long-term success.

[Andy Nogueira @andynog]

@brettcalhounn Exactly right, but many still don’t understand this. Concentration is not only subpar but also lacks resilience and anti-fragility. We need more decentralized businesses.

[Andy Nogueira @andynog]

While I agree that concentration and co-location can enhance value by creating quicker feedback loops, they don’t inherently make systems resilient or anti-fragile. Embracing diversification is essential to reduce risk effectively.

[Andy Nogueira @andynog]

@balajis And the best version of yourself is the one that empowers others to become the best version of themselves. Compounding effects.

[Andy Nogueira @andynog]

Exactly. In AI-native companies, it's not the model and the bare harness that form the moat. Instead, it's the customized workflow (harness) and the dataset (memory) that create the competitive edge. While the CPU (model plus harness) is a commodity, the real moat lies in the filesystem content, the memory itself.

[Andy Nogueira @andynog]

@adamshuaib Sorry, what grind? 😉👷

[Andy Nogueira @andynog]

Finding reality in a compressed model requires enormous energy and money, which is why only a few labs are developing foundation models. Additionally, data availability may be limited compared to real-world situations. Finding alpha by extracting the signal from noise requires depth, not breadth, to be economically viable. However, this process also requires proper simulations and validation.

[Andy Nogueira @andynog]

More global minima, fewer local minima, and higher efficiency lead to a better entropy distribution, encouraging greater and more effective diversification. In practical terms, survival is possible only for those who can detect a clearer signal amid the noise, as the solution space is explored more thoroughly through AI.

[Andy Nogueira @andynog]

It's always about the network, and I had a similar realization years ago when I read the Trillions book. Back then, IoT was the hype, AI was nonexistent, and Web3 was starting, but now they are all coming together. We've always had network effects, human and social, but now the number of nodes and connections has increased by 100 times or more, and the flow of entropy through it is much greater.
https://t.co/n1k8pAtP9N