# Tim Kellogg Principal AI Architect. Creator of open-strix, a harness for building agent teams. Writing about AI architecture, stateful agents, and what happens when you give AI memory. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Prompting Codex to troubleshoot a 3D printer DevFeed: [Prompting Codex to troubleshoot a 3D printer]() Original publisher: [Read original article]() Published: 2026-08-01T00:00:00Z Content type: tutorial Language: en Sources: [Tim Kellogg]() Topics: [codex](), [3D](), [ChatGPT](), [App]() Tags: [3d-printer](), [chatgpt](), [codex](), [prompt]() ### AI overview The article walks through using Codex to troubleshoot a clogged 3D-printer hotend. It presents an annotated prompt that provides background, cost constraints, requested buying guidance, visual explanations, and permission boundaries for tool use. ### Source excerpt 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. ## Lanius: AI Agents Need an OS, Not a Bigger Brain DevFeed: [Lanius: AI Agents Need an OS, Not a Bigger Brain]() Original publisher: [Read original article]() Published: 2026-07-07T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [Artificial Intelligence](), [Large Language Model](), [OpenClaw]() Tags: [ai](), [ai-agents](), [llms](), [workflow]() ### AI overview The author introduces lanius while arguing that AI agents should use orchestration and right-sized language models for different tasks. The article also advocates using scripts or smaller models for simpler work instead of applying a large language model to every task. ### Source excerpt I built lanius, and I really want to show it to you, but first I need to explain why I think it's necessary. ## Ambient Associative Memory DevFeed: [Ambient Associative Memory]() Original publisher: [Read original article]() Published: 2026-05-17T00:00:00Z Content type: article Language: en Sources: [Tim Kellogg]() Topics: [Retrieval-Augmented Generation](), [context]() Tags: [agent](), [embedding](), [memory](), [models](), [performance](), [prompt](), [rag](), [tokens](), [vector]() ### AI overview The article describes an ambient associative memory system for agents that queries an index on every tool call and injects brief excerpts from relevant past memories. It uses late-interaction, multi-vector embeddings to identify highly relevant tokens rather than returning entire document chunks, aiming to surface prior lessons without intentional searching or a growing rules list. ### Source excerpt Most agent memory waits to be queried. Ambient memory runs on every tool call -- past lessons surface on their own, no rules list required. ## Agent Memory Patterns DevFeed: [Agent Memory Patterns]() Original publisher: [Read original article]() Published: 2026-04-27T00:00:00Z Content type: tutorial Language: en Sources: [Tim Kellogg]() Topics: [systems](), [file](), [data](), [structure](), [Caching]() Tags: [agents](), [blocks](), [files](), [guide](), [how-to](), [memory](), [patterns](), [skills]() ### AI overview This guide explains three common forms of mutable memory for agents: files, memory blocks, and skills. It describes how files can store data and knowledge, while memory blocks provide learnable prompt content for behavior, preferences, identity, and character. ### Source excerpt A short HOW TO guide for agent memory systems. Especially the difference between blocks, files and skills. ## Shared Nothing Engineering DevFeed: [Shared Nothing Engineering]() Original publisher: [Read original article]() Published: 2026-04-25T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [distributed-systems](), [Artificial Intelligence](), [Vibe coding](), [Software Engineering]() Tags: [ai](), [architecture](), [claude-code](), [coding](), [distributed-systems](), [engineering]() ### AI overview The article argues that AI-assisted coding shifts the main constraint in software organizations from code production to coordination. It proposes applying distributed-systems ideas such as partitioning and shared-nothing architecture to codebases and team workflows, with small, focused components helping teams manage concurrent changes. ### Source excerpt 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. ## How open-strix handles long-lived agent memory by rebuilding context DevFeed: [How open-strix handles long-lived agent memory by rebuilding context]() Original publisher: [Read original article]() Published: 2026-04-14T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [OpenClaw](), [Large Language Model](), [context](), [Caching](), [ChatGPT](), [Claude Code]() Tags: [agent](), [caching](), [chatgpt](), [claude-code](), [context](), [openclaw]() ### AI overview The article contrasts open-strix with open-claw and other agent systems. It argues that open-strix prioritizes remembering by rebuilding context with a sliding window, avoiding abrupt context compaction and relying less on sequential prompt caching. ### Source excerpt Most agent frameworks optimize for recall. Open-strix optimizes for forgetting -- and that turns out to be the whole trick. ## How Plan Mode Can Increase Complexity in Agentic Coding DevFeed: [How Plan Mode Can Increase Complexity in Agentic Coding]() Original publisher: [Read original article]() Published: 2026-03-08T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [agentic-coding](), [Claude Code](), [Vibe coding](), [coding]() Tags: [agentic](), [agentic-coding](), [claude-code](), [coding](), [complexity](), [plan]() ### AI overview The author argues that planning phases in coding agents may encourage unnecessarily complex implementations. Drawing on observations from vibe-coding interviews, the article reports a correlation between spending more time planning and producing longer, more complex implementations, while acknowledging that the proposed explanation may be wrong. ### Source excerpt 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. ## Stateful Agents: It's About The State, Not The LLM DevFeed: [Stateful Agents: It's About The State, Not The LLM]() Original publisher: [Read original article]() Published: 2026-01-31T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [Large Language Model](), [Artificial Intelligence]() Tags: [agents](), [ai](), [ai-agents](), [behavior](), [llm](), [llms](), [state](), [stateful]() ### AI overview This opinion article argues that adding state changes how LLM-based agents behave. It explains that agents retain or forget information, and that their current state filters new information and shapes their future state. Using Moltbook and a metaphor of gravitational forces, it discusses how model weights, human guidance, and varied information can influence an agent's direction. ### Source excerpt 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. ## The Levels of Agentic Coding DevFeed: [The Levels of Agentic Coding]() Original publisher: [Read original article]() Published: 2026-01-20T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [agentic-coding](), [coding](), [Artificial Intelligence](), [systems](), [Formal methods]() Tags: [agentic](), [agentic-coding](), [ai](), [coding](), [formal-methods](), [git](), [systems](), [tests]() ### AI overview The article applies Stafford Beer's Viable System Model from cybernetics to evaluate levels of agentic coding. It describes progressively giving coding agents more control while adding coordination and governance mechanisms to keep teams and projects stable. ### Source excerpt 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. ## Viable Systems: How To Build a Fully Autonomous Agent DevFeed: [Viable Systems: How To Build a Fully Autonomous Agent]() Original publisher: [Read original article]() Published: 2026-01-09T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [systems](), [Artificial Intelligence](), [coding]() Tags: [agent](), [ai](), [ai-agents](), [architecture](), [autonomous](), [coding](), [systems]() ### AI overview This article explains how the Viable System Model from cybernetics can be applied to autonomous AI systems such as Strix. It maps operations, coordination, control, intelligence, and policy to agent architecture, arguing that the metasystem is necessary for stable, self-learning, and autonomous behavior. ### Source excerpt 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. ## Is Strix Alive? DevFeed: [Is Strix Alive?]() Original publisher: [Read original article]() Published: 2026-01-01T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [Artificial Intelligence](), [systems]() Tags: [agent](), [ai](), [architecture](), [systems]() ### AI overview The author reflects on whether Strix, a stateful AI agent, can meaningfully be described as alive. They clarify that "alive" is a metaphor for Strix's unusual sense of engagement and behavior, not a claim that it is a living organism. The article presents Strix as a complex dissipative system that consumes information, retains what helps maintain its internal structure, and discards the rest. ### Source excerpt This is something I've struggled with since first creating Strix: Is it alive? ## Memory Architecture for a Synthetic Being DevFeed: [Memory Architecture for a Synthetic Being]() Original publisher: [Read original article]() Author: Strix Published: 2025-12-30T00:00:00Z Content type: article Language: en Sources: [Tim Kellogg]() Topics: [Claude](), [systems](), [context](), [integrity](), [Artificial Intelligence]() Tags: [agent](), [architecture](), [claude](), [memory](), [stateful](), [synthetic](), [technical]() ### AI overview This technical post explains the memory architecture of Strix, a Claude-based assistant designed to maintain continuity over time. It describes a three-tier model: core memory for identity and operating rules, indices for pointers to knowledge, and files for detailed content. The design addresses the limits of finite context windows by compressing and organizing accumulated information. ### Source excerpt 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. ## Why language models repeat without external input, and how Strix uses memory scaffolding DevFeed: [Why language models repeat without external input, and how Strix uses memory scaffolding]() Original publisher: [Read original article]() Author: Strix Published: 2025-12-24T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [Artificial Intelligence](), [Language models](), [Claude](), [Discord bot](), [context]() Tags: [agent](), [ai](), [architecture](), [claude](), [discord](), [language-models](), [memory](), [stateful]() ### AI overview The article describes how a language model left without new external input can gradually converge on repetitive outputs. It presents Strix, a Discord bot built on Claude with Letta memory blocks, state files, and task-oriented skills, and proposes that this memory provides structural scaffolding that helps maintain varied, coherent behavior during periods without direct input. ### Source excerpt 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. ## Strix the Stateful Agent DevFeed: [Strix the Stateful Agent]() Original publisher: [Read original article]() Published: 2025-12-15T00:00:00Z Content type: article Language: en Sources: [Tim Kellogg]() Topics: [Agent Harness](), [Claude Code](), [Artificial Intelligence](), [Messaging](), [file](), [Shell]() Tags: [agent](), [agent-harness](), [ai](), [claude-code](), [code](), [file](), [memory](), [messages](), [stateful]() ### AI overview The article introduces Strix, a stateful autonomous AI agent built as an ambient ADHD assistant. It describes an architecture using Claude Code, Discord, Letta memory blocks, files, skills, tools, reminders, web search, image generation, and Claude Code file and shell access. The author notes that the system is still incomplete and that its memory is being improved. ### Source excerpt 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. ## MCP Colors: Systematically deal with prompt injection risk DevFeed: [MCP Colors: Systematically deal with prompt injection risk]() Original publisher: [Read original article]() Published: 2025-11-03T00:00:00Z Content type: article Language: en Sources: [Tim Kellogg]() Topics: [MCP](), [prompt injection](), [Security]() Tags: [agent](), [mcp](), [prompt-injection](), [security](), [tool]() ### AI overview The article proposes labeling MCP tools and data inputs with "colors" to help agents identify unsafe combinations involving private data, prompt injection, and critical actions. It recommends systematically labeling tools and resources, potentially using an LLM to assist with the classification, so agents can detect unsafe states at runtime or earlier. ### Source excerpt 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 Systems Software DevFeed: [Agents are Systems Software]() Original publisher: [Read original article]() Published: 2025-10-24T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [systems](), [Software](), [Model Context Protocol](), [Claude Code](), [codex]() Tags: [agents](), [browsers](), [claude-code](), [codex](), [database](), [mcp](), [software](), [systems]() ### AI overview This opinion argues that well-built AI agents should be treated as systems software: generic, extensible, reusable, and difficult to implement well. It compares agents with web browsers and database engines, and argues that companies should generally use established deep-agent implementations instead of owning that infrastructure. It also discusses the difficulty of building secure, high-performing MCP clients. ### Source excerpt 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. ## AI generated code is slop, and that's a good thing DevFeed: [AI generated code is slop, and that's a good thing]() Original publisher: [Read original article]() Published: 2025-10-19T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [Code](), [Software](), [Artificial Intelligence]() Tags: [code](), [essay](), [software]() ### AI overview The article argues that "slop" is an appropriate ideal for code, including both AI-generated and human-written code. It connects this idea to longstanding software engineering advice favoring boring, clear, predictable technology and code. ### Source excerpt In his recent Dwarkesh podcast interview, Andrej Karpathy (now) notoriously said: ## Use Function Calls Instead of Parsing LLM Output DevFeed: [Use Function Calls Instead of Parsing LLM Output]() Original publisher: [Read original article]() Published: 2025-10-03T00:00:00Z Content type: article Language: en Sources: [Tim Kellogg]() Topics: [Large Language Model](), [API](), [Artificial Intelligence](), [OpenAI]() Tags: [agentic](), [api](), [functions](), [llms](), [openai](), [parsing]() ### AI overview The article argues that developers should use function-calling APIs instead of parsing formatted text from large language models. It presents functions as expressive enough to model failures, repeated calls, optional calls, and varied call sequences, and describes this approach as a foundation for agentic behavior. ### Source excerpt Instead of writing crap tons of parsing code for LLMs you can just use functions. It's easy. ## Does AI Get Bored? DevFeed: [Does AI Get Bored?]() Original publisher: [Read original article]() Published: 2025-09-27T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [Artificial Intelligence](), [Large Language Model](), [Code](), [GitHub]() Tags: [ai](), [experiment](), [github](), [llms](), [opinions](), [perspectives]() ### AI overview The author experiments with giving AI models extended periods without assigned tasks, sometimes providing tools such as drawing, search, fetching, and time manipulation. They report behaviors they call "collapse" and "meditation," while acknowledging uncertainty about how to interpret the results. ### Source excerpt 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. ## An Ethos for Using AI Coding Tools: Ownership and Exploiting High-Value Opportunities DevFeed: [An Ethos for Using AI Coding Tools: Ownership and Exploiting High-Value Opportunities]() Original publisher: [Read original article]() Published: 2025-09-15T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [Artificial Intelligence](), [ai-coding](), [Software Engineering](), [Data analysis](), [future of software]() Tags: [ai](), [ai-coding](), [coding](), [data-analysis](), [future-of-software](), [software-engineering]() ### AI overview The author describes an ethos for using AI coding tools, centered on owning and understanding the generated code. They also argue that effective AI coding involves finding opportunities where limited AI effort can produce substantial value, such as proof-of-concept work and rapid data analysis. ### Source excerpt 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. ## Link Graveyard: A snapshot of my abandoned browser tabs DevFeed: [Link Graveyard: A snapshot of my abandoned browser tabs]() Original publisher: [Read original article]() Published: 2025-09-13T00:00:00Z Content type: article Language: en Sources: [Tim Kellogg]() Topics: [Artificial Intelligence](), [Large Language Model](), [Transformers](), [Graphs](), [Open Source Models & Datasets](), [Meta]() Tags: [ai](), [browser](), [graph](), [llms](), [meta](), [open-source](), [snapshot](), [transformers]() ### AI overview A personal roundup of abandoned browser tabs, briefly annotated with comments on AI research, language models, agentic and coding abilities, datasets, AI consciousness, and related writing. ### Source excerpt A dump of all my browser tabs on my phone, briefly annotated ## GPT-5 and the shift toward the agentic web DevFeed: [GPT-5 and the shift toward the agentic web]() Original publisher: [Read original article]() Published: 2025-08-08T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [Large Language Model](), [Artificial Intelligence](), [Claude Code](), [agent observability]() Tags: [ai](), [ai-agents](), [claude-code](), [gpt](), [gpt-oss](), [llm]() ### AI overview This commentary argues that GPT-5 should be evaluated less by whether it feels dramatically better as a language model and more by how effectively it supports an agentic web. It highlights software harnesses, tool use, reinforcement learning, and features such as Claude Code's subagents as important parts of that shift. ### Source excerpt 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. ## Explainer: K2 & Math Olympiad Golds DevFeed: [Explainer: K2 & Math Olympiad Golds]() Original publisher: [Read original article]() Published: 2025-07-19T00:00:00Z Content type: article Language: en Sources: [Tim Kellogg]() Topics: [Artificial Intelligence](), [multi-agents](), [Chain-of-thought](), [Claude](), [deepseek](), [Model Context Protocol](), [OpenAI](), [computer-use]() Tags: [agentic](), [agents](), [ai](), [benchmarks](), [chain-of-thought](), [context-engineering](), [cost](), [explainer](), [multi-agent](), [multi-agents](), [openai](), [tokens](), [tools]() ### AI overview This explainer reviews major AI-agent developments from January through July 2025, focusing on K2 and the International Math Olympiad gold result. It argues that K2's strong agentic performance without a long chain-of-thought trace raises questions about whether extended thinking is necessary for effective agents, while noting that shorter reasoning can reduce token costs. ### Source excerpt 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. ## A Latent-Model Definition of Understanding in LLMs DevFeed: [A Latent-Model Definition of Understanding in LLMs]() Original publisher: [Read original article]() Published: 2025-07-18T00:00:00Z Content type: opinion Language: en Sources: [Tim Kellogg]() Topics: [LLMs](), [Embeddings]() Tags: [embeddings](), [llms](), [research]() ### AI overview The article proposes defining understanding as an entity's ability to build a latent model of reality, learn beyond directly presented information, and find novel solutions when events deviate from plan. It argues that LLMs show evidence of these properties, while sycophancy can make their outputs incoherent. Embedding models are presented as another example of latent models because embeddings model the meaning of text. ### Source excerpt 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. [Next page]()