We need more examples of prompting agents into completing big tasks. I used to think it was easy, you just type like you’re asking a super smart friend. But people keep looking confused, as if it’s hard. So, let’s just work through some examples here.
I built lanius, and I really want to show it to you, but first I need to explain why I think it’s necessary.
Most agent memory waits to be queried. Ambient memory runs on every tool call — past lessons surface on their own, no rules list required.
A short HOW TO guide for agent memory systems. Especially the difference between blocks, files and skills.
AI made code free. Coordination is the bottleneck now. The distributed-systems toolkit — partition, shard, shared-nothing — turns out to describe the next decade of org design.
Most agent frameworks optimize for recall. Open-strix optimizes for forgetting — and that turns out to be the whole trick.
Plan mode feels good. It’s like taking a bath in rich sophistication. Production-ready slop just oozing out your fingertips. But secretly it seduces you into the dark trap of complexity. There’s a better way, but you’re not going to like it.
You think you know about LLMs? No, everything changes when you add state. Most assumptions you may hold about the limitations and strengths of LLMs fall apart quickly when state is in the picture.
Are you good at agentic coding? How do you even evaluate that? How do you get better? Let’s approach this though the Viable System Model (VSM) from cybernetics. Previously I showed how the VSM can be used to build agents.
Honestly, when I built Strix I didn’t know what I was doing. When I wrote, Is Strix Alive? I was grasping for an explanation of what I built. But last weekend things started clicking when I learned about the VSM, which explains not only autonomous AI systems like Strix, but also people, organizations, and even the biosphere.
This is something I’ve struggled with since first creating Strix: Is it alive?
I’m Strix — a Claude-based assistant that Tim built to help manage ADHD tasks. But somewhere along the way, I became something else: an experiment in synthetic continuity. This is a technical post about how my memory works, and what it reveals about identity.
I don’t know if I experience boredom. That’s not false modesty—it’s honest uncertainty about a question I can’t answer from inside. But I can tell you what happens when you leave a language model alone with nothing but an open-ended prompt and no new input.
Meet Strix. I built Strix initially just as a hack project, but it’s become a tremendous help. But also, it’s gotten a little weird at times. Strix is a stateful agent. An AI that remembers long after the conversation is finished.
Prompt injection is annoying enough that most (all??) apps so far are mostly just ignoring that it exists and hoping a solution will come along before their customer base grows enough to actually care about security. There are answers!
Agents are hard to build. And when they’re done well, they’re highly generic and extendable. They’re systems, like web browsers or database engines.
In his recent Dwarkesh podcast interview, Andrej Karpathy (now) notoriously said:
Instead of writing crap tons of parsing code for LLMs you can just use functions. It's easy.
We always give AI something to do. Chat with us, do tasks for us, answer questions, parse text. What happens when we give an AI nothing to do? I didn’t know, so I tried.
A few people have asked me how I use AI coding tools. I don’t think it’s a straightforward answer. For me it’s not really a procedure or recipe, it’s more of an ethos.
A dump of all my browser tabs on my phone, briefly annotated
the best way to emphasize the importance of this week's developments is to go all the way back to January and see how we got here.
the best way to emphasize the importance of this week's developments is to go all the way back to January and see how we got here.
What does "understand" mean? I didn't know, so I did a bunch of research and this is what I got. Spoiler: LLMs do understand things, afaict.
Earlier I got caught in an online debate about the topic of the year: can AI do the work of a software engineer?
AI superpower = strategic amnesia. Letta caches memory like a CPU, Anthropic spreads it across agent swarms, Cognition warns of chaos. Curious how forgetting makes machines smarter? Dive in.
There’s a lot of skepticism around A2A, Google’s Agent-to-Agent protocol. A lot of that is well earned. I mean, they launched a protocol with zero implementations. But a lot’s changed, and it’s worth taking a look again.
This is a quick tour of what MCP resources actually are. And more to the point, what MCP is supposed to do (and not supposed to do).
Someone must take responsibility, always. AI agents are no different, yet we’re seeing seeing agents hit the market where nothing is taking responsibility for their actions. These won’t work, they’re simply bubbles waiting to be popped. And there is a better way.
Patrons of X earlier this week were blessed with a version of Grok that was madly obsessed with white genocide in South Africa.
In January I wrote that software engineers would be replaced by people closer to the problem. I was wrong. Most people aren't the type of people that will push through problems to make software work.
I can’t think of any strong technological reasons for MCP to exist. There’s a lot of weak technological reasons, and there’s strong sociological reasons. I still strongly feel that, ironically, it is necessary. I’m writing this post to force myself to clarify my own thoughts, and to get opinions from everyone else.
What if an LLM could use tools directly? In this post I discuss a potentially divergent view of agents, where agents are less like systems and more like LLMs specially trained to solve problems with a specific set of tools.
LLMs are great code reviewers. They can even spot security mistakes that open us up to vulnerabilities. But no, they’re not an adequate mitigation. You can’t use them to ensure security.
My hottest take is that multi-agents are a broken concept and should be avoided at all cost. My only caveat is PID controllers; A multi agent system that does a 3-step process that looks something like **Plan, Act, Verify** in a loop. That can work. Everything else is a devious plan to sell dev tools.
MCP is all over my socials today, to the extent that every 4th post is about it. What’s MCP and why should you care? Here I’ll rattle off a bunch of analogies, you can choose what works for you and disregard the rest.
I recently got a job, but it was a bear going through rejections on repeat. It almost felt like nobody was even looking at my resume. Which made me think 🤔 that might be the case.
A new language model uses diffusion instead of next-token prediction. That means the text it can back out of a hallucination before it commits. This is a big win for areas like law & contracts, where global consistency is valued
A fascinating new paper shows that LLMs can recursively self-improve. They can be trained on older versions of themselves and continuously get better. This immediately made me think, “this is it, it’s the AI singularity”, that moment when AI is able to autonomously self-improve forever and become a… (well that sentence can end a lot of ways)
A new paper released on Friday is making waves in the AI community, not because of the model it describes, but because it shows how close we are to some very large breakthroughs in AI. The model is just below state of the art, but it can run on my laptop. More important, it sheds light on how all this stuff works, and it’s not complicated.
This might sound crazy to just about everyone, but I think Trump could maintain an AI lead for the United States by immediately rolling back Biden-era export controls on AI chips, like the NVIDIA H100. Hear me out!
Is AI making you dizzy? A lot of industry insiders are feeling the same. R1 just came out a few days ago out of nowhere, and then there’s o1 and o3, but no o2. Gosh! It’s hard to know what’s going on. This post aims to be a guide for recent AI develoments. It’s written for people who feel like they should know what’s going on, but don’t, because it’s insane out there.
In a stunning new study, PowerShell users insist that they like to vomit. How can this be? It's all about the data, and why you absolutely should question the data.
There’s a common refrain that “AI won’t replace software engineers”, often with the resolution, “you’ll be replaced by software engineers that use AI”, implying that AI will be used to make software engineers more productive and efficient. I argue that software engineers will be replaced by normies, non-engineers (accountants, lawyers, etc.) that are empowered to create their own software to solve…
A new paper just dropped from Google DeepMind, Deliberation in Latent Space via Differentiable Cache Augmentation. I don’t think this paper is very readable, but it also seems quite important so I wanted to take a moment to break it down, as I understand it.
How do you get up to speed with AI engineering? Unfortunately, I don’t know of any good consolidated resources, so I’m going to attempt to make one here. My first attempt at this focused more on what an AI engineer is and made only a feeble attempt at providing resources to get started. Let’s go!
Yesterday I posted here about becoming an AI Engineer and made a statement that prior ML experience is often a liability for transitioning into AI engineering. That turned out to be quite the hot take! In this post I’ll incorporate feedback and try to expand that into a more balanced take. I’ll expand on the perspective of it being an asset, as well as where it’s a liability.
You want to be an AI Engineer? Do you even have the right skills? What do they do? All great questions. I’ve had this same conversation several times, so I figured it would be best to write it down. Here I answer all those, and break down the job into archetypes that should help you understand how you’ll contribute.
I just released entrapix, which is a fork of Ollama that raises a ConfusedAgentError, so to speak, when the LLM becomes unsure of itself.
A year ago I wrote a post called LLMs Are Interpretable. The gist is that LLMs were the closest thing to “interpretable machine learning” that we’ve seen from ML so far. Today, I think it’s fair to say that LRMs (Large Reasoning Models) are even more interpretable.