Posts

When you enable extended thinking on Claude or switch to an o-series model, the price per request jumps 3 to 10x. That is not because you are getting a smarter model. You are getting the same model spending more computation on your specific question at inference time. Understanding the difference between training-time compute and inference-time compute changes how you decide when these models are worth using. How a Standard Model Generates Output A standard LLM takes all the input tokens (system prompt, history, user message), runs them through the transformer layers in a single forward pass to predict the next token, then samples that token and appends it to the sequence. It repeats this process until it generates an end token. Each forward pass scales with model size. Total cost scales linearly with output token count. ...

Most engineers building agents spend time worrying about hallucinations. The more immediate risk is that your agent will faithfully execute instructions planted by an attacker in a web page, a retrieved document, or an MCP server response. No jailbreak required. The model follows the injected instructions because to the model, they look exactly like legitimate instructions. This is not a theoretical risk. Researchers have demonstrated successful attacks against agents with web search tools, email agents that read malicious messages, and code agents that load poisoned documentation. As agents gain more autonomy - file access, email, database writes, API calls - the blast radius of a successful injection grows proportionally. ...

The dirty secret of most AI product teams in 2026: when someone asks “how do you know the new prompt is better?” the honest answer is “we ran it on a few examples and it felt better.” That is vibes-based QA. It works for a demo and collapses in production. LLM features break in ways that unit tests do not catch. You change the prompt to fix one user’s complaint, and it silently regresses the output for 30 other inputs you never checked. You upgrade the model version and the tone shifts. You add a retrieved document to the context and the model stops following the output schema. None of these have stack traces. The failure arrives as a support ticket three days later. ...

Ask anyone building agents what their biggest problem is and “memory” comes up within the first two sentences. The agent solved the bug yesterday and has no idea today. It learned the user prefers TypeScript, then suggested JavaScript an hour later. It re-read the same 4,000-line file three times in one session because nothing told it that it already knew the answer. The confusing part is that everyone thinks they already have memory. “I have a 1M token context window.” That is not memory. That is a desk. Memory is the filing cabinet you reach for when the desk is full and the meeting ended. This post is about how agents actually remember in 2026 - starting from the naive version everyone builds first, showing exactly where it breaks, and ending at the hybrid architecture that production agents converge on. ...

The term “vibe coding” entered the vocabulary in early 2025. The idea: describe what you want, let the AI write it, accept the output, repeat. No architecture. No review. No tests. Just vibes. It works surprisingly well for prototypes and one-off scripts. It fails catastrophically for anything that needs to survive contact with production. AI-first development is the opposite approach applied to the same tools. It uses AI as a force multiplier inside a disciplined engineering workflow - not as a replacement for engineering judgment. The distinction matters because the tools are identical. The outcomes are not. ...

Every AI framework promises agents that can “autonomously complete complex tasks.” The demo shows an agent booking flights, writing code, and sending emails. Then you deploy it and it hallucinates a nonexistent API endpoint, gets stuck in an infinite loop calling the same tool, and racks up $200 in API costs before your rate limiter kicks in. I have shipped AI agents to production across three different products in the last year. Here is what actually works, what fails, and the patterns that separate a demo from a system your users can rely on. ...

Claude Code has a plugin system that requires no code. Skills are markdown files that live in ~/.claude/skills/ and define reusable workflows, checklists, and behaviors that Claude can invoke during sessions. They are the simplest way to extend Claude Code’s capabilities and the most underused feature in the toolset. What Skills Are A skill is a markdown file that describes a workflow or set of instructions. When a skill is relevant to the current task, Claude loads it and follows the instructions. Skills can also be invoked explicitly as slash commands. ...

Claude Code has two mechanisms for controlling behavior: CLAUDE.md instructions and hooks. They solve different problems. CLAUDE.md is guidance - Claude reads it and mostly follows it. Hooks are executable scripts that fire on specific events - they run deterministically, every time, regardless of what Claude decides to do. Understanding the boundary between these two systems is the difference between a setup that mostly works and one that always works. ...

The serverless landscape has shifted. AWS Lambda defined the category, but Cloudflare Workers has been quietly eating into Lambda’s territory with a fundamentally different execution model. If you are evaluating these two platforms in 2026, the decision is no longer obvious. Here is a technical breakdown of what matters and when each platform wins. Execution Model - V8 Isolates vs Containers This is the most important architectural difference and it drives everything else. ...

Every team building with LLMs eventually hits the same question: should we fine-tune, improve our prompts, or add retrieval? The answer in 2026 is almost never just one of these. But knowing which lever to pull first - and when to combine them - is the difference between a system that works and one that burns money while hallucinating. The Three Approaches - Defined Precisely Prompting means controlling the model’s behavior through the input text alone. System prompts, few-shot examples, chain-of-thought instructions, structured output schemas. The model’s weights do not change. You are working within its existing capabilities. ...

The “prompt engineer” job title had a good run. From late 2023 through mid-2025, companies hired people whose primary skill was writing effective prompts for language models. Job postings asked for “experience crafting system prompts” and “prompt optimization techniques.” Some of these roles paid remarkably well for what amounted to writing structured English. By 2026, the role has evolved - or more accurately, split. Simple prompt crafting got absorbed into every developer’s toolkit. The complex work evolved into AI engineering: building systems where language models are components, not standalone tools. The difference between a prompt engineer and an AI engineer is the difference between writing a SQL query and designing a database architecture. Both involve the same technology. One is a skill. The other is a discipline. ...

The API protocol question is not “which is best” but “which is best for this specific communication pattern.” After years of teams making this choice wrong - migrating to GraphQL for internal services, using REST for high-throughput microservices, or adopting gRPC for public APIs - the decision criteria are now clear. Here is the decision tree, backed by real performance data and hard-won lessons. The 30-Second Decision Internal service-to-service, high throughput - gRPC Public-facing API consumed by third parties - REST Client applications with varied data needs - GraphQL Simple CRUD with minimal clients - REST If your use case maps clearly to one of these, you probably do not need to read further. If it does not, the details below will help. ...

Bun 2.0 is not just faster than Node.js. It ships a bundler, test runner, package manager, and native TypeScript support in a single binary. Here is what actually matters and when to switch.

I have been using Claude 4 for production coding work since its launch. Not as a toy, not for demos - for actual engineering tasks: refactoring a 50,000-line codebase, building new features from scratch, debugging race conditions, and writing infrastructure code. This is a technical assessment of what has changed, what genuinely works, and where it still falls short. Extended Thinking - The Feature That Changed Everything Previous models generated code by predicting the next token. Claude 4 with extended thinking actually reasons about problems before writing code. The difference is not subtle. ...

Claude Code is not autocomplete with a chat window. It is an agentic coding assistant that reads your codebase, runs commands, edits files, and executes multi-step tasks from your terminal.

In Q4 2025, several high-profile companies publicly announced they were “reducing engineering headcount by 30-50% thanks to AI.” Six months later, most of them are quietly rehiring. The pattern is consistent enough to draw conclusions from. This is not an argument that AI is not useful for software engineering. It clearly is. But the framing of “replacement” versus “multiplier” makes all the difference between a successful AI adoption and an expensive disaster. ...

Traditional VPNs are miserable. They funnel all traffic through a central gateway, add latency to every connection, require manual configuration, and break constantly when network conditions change. Tailscale reimagined the entire model by building a mesh network on top of WireGuard, and the result is a VPN that feels invisible. Here is how it works under the hood and why the architecture is worth understanding even if you never use Tailscale. ...

Claude Code bills by tokens. Every file read, every tool schema, every message in the conversation - it all goes into the context window, and the context window determines cost. Most developers treat context as unlimited and invisible. It is neither. Understanding how context accumulates and managing it deliberately can cut Claude Code costs by 70% or more without reducing output quality. How Context Accumulates A fresh Claude Code session starts with a base context: system prompt, CLAUDE.md content, MCP tool schemas, and the initial user message. A typical starting context is 3,000-8,000 tokens depending on configuration. ...

AI-generated code has a trust problem. Not because it is always wrong - it is right often enough to be dangerous. The failure mode is not obvious errors that fail to compile. It is subtle mistakes that pass tests, look reasonable in review, and break in production three weeks later. Reviewing AI output effectively requires understanding what AI gets wrong consistently and building a review process that catches those specific failure patterns. ...

Most teams using AI in production make the same mistake: they pick one model and use it for everything. Every autocomplete suggestion, every code generation task, every architecture review goes through the same frontier model. This is like hiring a senior architect to fix typos. It works, but the cost-to-value ratio is absurd. Model routing is the practice of directing each task to the model best suited for it - matching task complexity to model capability. Done right, it cuts API costs by 5-10x while maintaining or improving output quality, because smaller models are often better at simple tasks than large models that overthink them. ...