The better model is not just a little better. In daily coding work, it can save huge amounts of time and mental energy. The difference is often whether you spend an hour going back and forth with the model or simply get the problem solved. I recently experienced this with GLM-5.3 Flash. I had been … Continue reading Use Better Model; Fable5.1 > GLM-5.3 Flash

The essence of an agent is surprisingly simple: codify the know-how of an expert. AI is already remarkably capable—it can reason, write code, analyze data, and solve many difficult problems. What it often lacks is the specific workflow and practical know-how of a particular job. It doesn't automatically know all the details, nuances, conventions, and … Continue reading The Essence of an Agent

The older I get, the more I believe that the ability to think clearly—to understand how the world works, how people behave, and what truly matters—is the foundation of a happy, fulfilling, rich, and successful life. Much of what I once experienced as mental or spiritual pain now seems, at least in part, to come … Continue reading Thinkability: Learning How the World Works

There are broadly two types of people when it comes to someone like Donald Trump. Trump constantly promotes himself. He doesn't seem particularly embarrassed by self-promotion. He can say, in effect: "I'm the best. I'm successful. I'm great." Most people, at least internally, would probably have a reaction like: That sounds very arrogant. How could … Continue reading Why Some People Love the Trump Type and Others Can’t Stand It

Context degradation is real. This session started with a long summary of prior work which is useful for continuity but it also means: stale information gets treated as current, AI check memory which could be outdated. The more context I carry, the harder it is to distinguish what's verified vs what's just noted. lose track of user's question … Continue reading AI Degradation in Long Contextual Conversations

Every session starts stateless; the MCP server provides excellent tools, but the real domain expert agent needs to know how to use these tools proficiently and intelligently. Additionally, they must be conscious of token consumption, as it is not abundant and time is critical. I don't want to waste a lot of time and money … Continue reading The Approach to Build Domain Expert Agent

An artifact is a persistent, addressable object—such as a dataset, file, report, or HTML output—stored outside the LLM's conversation context and referenced by an ID. When creating agents for extensive data-heavy tasks, it quickly becomes apparent that the large amounts of data displayed in the conversation field are unwieldy and unnecessary. Users only need to … Continue reading Use Artifacts in the Agents

I have an idea to let MedGemma become the specialized local engine, MCP becomes the interface, and Claude/ChatGPT becomes the high-level reasoning/orchestration layer. The flow would be Claude Desktop sends a tool call → the MCP server builds a prompt → llama-server runs MedGemma inference on CPU → structured JSON comes back to Claude. Deploying … Continue reading Deploying MedGamma Mini Model Locally and Create an MCP Server

1. Standard OAuth 2.1 (MCP spec-compliant) Who uses it: Remote/hosted MCP servers (e.g., a server on Fly.io, Cloudflare, etc.) How: Full OAuth 2.1 flow — the MCP server is a resource server, a separate auth server (Auth0, Okta, Keycloak, etc.) issues tokens. Client discovers auth server via /.well-known/oauth-protected-resource, opens browser with PKCE, exchanges authorization code for access token. Example: An MCP … Continue reading The 4 Common OAuth/auth Patterns for MCP Servers

An AI agent can reason, write code, and call tools. But the web is where most real-world information lives. So how does an agent actually use a web page? The answer has evolved through three distinct generations, and which one your agent uses determines what it can and can't do. Generation 1: Fetch and Parse The simplest approach. The agent makes an HTTP request, … Continue reading How AI Agents Browse the Web: From Fetch to Click