by armin-ronacher
Coding agents dominate reinforcement learning data, which means the smartest way to build any agent is to lean into the coding agent paradigm — files, code execution, Unix workflows — rather than fight it with custom tools.
by pewdiepie
PewDiePie fine-tunes a Qwen 32B model into a coding specialist that scores 39.1% on Aider Polyglot — beating older ChatGPT and Gemini versions — while documenting every failure, melted cable, and dead GPU along the way.
by amp
The IDE sidebar is a dead-end interaction model for coding agents—parallel, headless agent swarms that run for 45 minutes without human input replace the one-on-one assistant workflow.
by nicholas-carlini
Sixteen Claude instances working in parallel without human supervision can produce a 100,000-line Rust-based C compiler capable of compiling the Linux kernel—but only when the task verifier and CI pipeline are nearly perfect.
by alex-shershebnev
Building AI coding agents requires only basic tooling—a ReAct loop, tool definitions, and MCP integration—and the developer's role shifts from writing code to managing autonomous virtual developers.
by michael-bolin
The agent loop—a simple cycle of LLM calls and tool execution—is the core of every AI agent, but performance requires stateless design, prompt caching, and context compaction to avoid quadratic inference costs.
by aicodeking
Kimi K2.5 challenges proprietary models at a fraction of the cost: trillion-parameter MoE with vision, agent swarm parallelism, and 5th place on AICodeKing's benchmark—beating Claude Sonnet 4.5 and DeepSeek V3.2.
by amp
The assistant era is over—agents now write production code. The next frontier is building 'agent-native codebases' with feedback loops that let agents verify their own work autonomously.
by adam-gospodarczyk
MCP's flaws matter less than the fundamental LLM limitations it exposes: context bloat from tool schemas, degraded instruction following in long conversations, and inference costs that balloon with agentic workflows.
by hacker-news-community
Simple retrieval often outperforms complex vector infrastructure—BM25, SQLite FTS5, and grep handle most local RAG use cases better than dedicated vector databases.
by alexander-opalic
Learn how to build a conversational AI that queries your personal knowledge base using Nuxt, Nuxt Content, and the Anthropic SDK.
by david-fant
You can extract and reconstruct any React component from a production website by leveraging React Fiber's internal tree structure combined with LLMs.
by mattpocockuk, dex-horthy
Live conversation exploring practical approaches to AI-assisted coding, context engineering, and building reliable agents in complex codebases.
by github
Build AI coding assistants by connecting to Copilot CLI through JSON-RPC, letting you embed GitHub's coding agent into any application.
by geoffrey-huntley
Coding agents are just 300 lines of code running in a loop—demystifying AI tooling reveals that the model does the heavy lifting, and understanding these primitives transforms you from AI consumer to AI producer.
by simon-willison
Claude Cowork repackages Claude Code's powerful agentic capabilities for general audiences through accessible design rather than technical innovation—a pragmatic approach to unlock untapped value.
by geoffrey-huntley
A hands-on Go workshop that builds a coding agent incrementally through six files, each adding one capability - proving agents need only simple primitives composed in a loop.
by simon-willison
Reasoning models have made LLM-generated code undeniably good, and 2026 will bring both a major security incident from coding agents and the resolution of the sandboxing problem.
by humanlayer-team
Systematic context management—through frequent intentional compaction and a Research-Plan-Implement workflow—enables productive AI-assisted development in complex production codebases.
by langchain
Context engineering—filling the context window with the right information at each step—determines agent performance more than model choice or complex frameworks.
by jediah-katz
Coding agents perform better when they pull context on demand rather than receiving everything upfront—files serve as a simple, future-proof abstraction for this dynamic retrieval.
by thariq-shihipar
Bash is the most powerful agent tool. The Claude Agent SDK packages Claude Code's battle-tested patterns—tools, file system, skills, sandboxing—for building coding and non-coding agents alike.
by philipp-schmid
As AI models converge in benchmark performance, the infrastructure managing them—Agent Harnesses—becomes the competitive differentiator for building reliable, multi-day workflows.
by gergely-orosz, chip-huyen
Chip Huyen explains how AI engineering differs from ML engineering, walking through the practical developmental path from prompts to RAG to fine-tuning.
by chip-huyen
A practitioner's guide to building applications on foundation models, covering prompt engineering, RAG, finetuning, agents, and evaluation.
by trey-grainger, doug-turnbull, max-irwin
The holy grail for AI-powered search lies at the intersection of semantic search, personalized search, and domain-aware recommendations—systems that understand the domain, the user, and can match arbitrary queries to any content.
by gergely-orosz, martin-fowler
Martin Fowler argues AI represents the biggest shift in software engineering since assembly to high-level languages—not because of abstraction level, but because we now work with non-deterministic systems.
by chip-huyen
A comprehensive guide explaining the three-phase process (pretraining, supervised fine-tuning, RLHF) used to train models like ChatGPT.
by stephen-wolfram
Explains how ChatGPT works by breaking down neural networks, embeddings, and training—connecting modern AI to foundational questions about language and thought.
by steven-bartlett
Tristan Harris warns that AI companies are racing to build a 'digital god' that could automate all human cognitive labor, with insiders believing this will happen within 2-10 years while publicly downplaying the risks.
by andrej-karpathy
A comprehensive introduction to how large language models work, covering the entire pipeline from internet data collection through tokenization, neural network training, and inference.
by mo-gawdat, steven-bartlett
Mo Gawdat, former Chief Business Officer at Google X, warns that AI represents humanity's greatest existential challenge—bigger than climate change—and outlines his 'three inevitables' for why we're approaching a point of no return.
by carl-assmann
CLI tool that enables LLM conversations through markdown files in your preferred editor, storing chat history as readable documents.
by dex-horthy
A practical framework for getting AI coding agents to work reliably in brownfield codebases through context engineering, intentional compaction, and the Research-Plan-Implement workflow.
by varin-nair
Frontier models cap out at 1-2 million tokens, yet enterprise codebases span several million. Factory's solution: a five-layer context stack that delivers the right information at the right time.
by antonio-gulli
A practical guide presenting 21 design patterns for building AI agents, covering prompt chaining, tool use, multi-agent collaboration, and self-correction techniques with examples in LangChain, CrewAI, and Google ADK.
Patterns and practices for building autonomous AI agents with prompt chaining, routing, and reflection
by andrej-karpathy, dwarkesh-patel
Karpathy argues we're building 'ghosts' that imitate internet documents rather than evolved animals, and that practical AI agents will take a decade—not a year—to fully mature.
by maham-codes
A practical guide to building AI agents using prompt chaining and basic primitives instead of heavy frameworks.
by mark-anthony-cianfrani
A first-principles guide to building AI agents in TypeScript, covering LLM integration, conversation memory, tool calling, and agentic loops.
by erik-schluntz, barry-zhang
Anthropic's guide to building agentic LLM systems, advocating for simple composable patterns over complex frameworks.
by andrej-karpathy
Andrej Karpathy's practical guide to using LLMs effectively: understanding them as lossy internet zip files, managing token windows, selecting models across providers, and leveraging thinking models for complex problems.
by mark-winteringham
Practical guide to using generative AI for test design, synthetic data generation, and automation without the hype.
by yacine-mahdid
Prompt chaining breaks complex LLM tasks into sequential steps, each with a specific job and structured input/output, trading latency for reliability.
by humanlayer-team
A manifesto for building production-grade LLM agents, arguing that effective agents combine mostly deterministic software with strategic LLM decision-making rather than naive 'loop until solved' patterns.
by simon-willison
Simon Willison's comprehensive year-in-review analyzing how reasoning models, autonomous agents, and Chinese AI competition fundamentally reshaped the landscape in 2025.
Techniques for providing AI agents and LLMs with optimized context
by humanlayer-team
Guidelines for crafting an effective CLAUDE.md file, emphasizing brevity, universal applicability, and progressive disclosure to maximize Claude Code's instruction-following capacity.
by andrew-qu
Stripping a text-to-SQL agent down to a single bash tool produced a 3.5x speedup, 100% success rate, and 37% fewer tokens—proving that simpler agent architectures outperform elaborate tooling.
by dennis-stanoev
Understanding AI agents requires building one from scratch: an agent is a wrapper around an LLM that makes decisions and takes actions through a simple tool-use loop.
by xiwei-xu
Context engineering—not model fine-tuning—should be the central challenge for generative AI systems, solved through a Unix-inspired file system abstraction that treats all context components uniformly.
by mario-zechner
Minimal coding agents outperform bloated ones because frontier models already understand agentic coding—so the harness should stay out of the way with fewer than 1,000 tokens of system prompt and just four tools.
by ryan-x-charles
Markdown has become a general-purpose programming language—AI agents like Claude Code compile structured specifications into working applications.
by anthropic
Context is a finite resource in LLM agents; treating tokens as precious budget rather than limitless capacity enables reliable long-horizon task completion.
by geoffrey-huntley
Coding agents require only 300 lines of code in a loop with LLM tokens - understanding these fundamentals transforms you from AI consumer to producer.
by thorsten-ball
Building a functional coding agent requires only an LLM, a loop, and a handful of tool definitions—the complexity lies in refinement, not architecture.