Articles
How to Deploy an AI Agent
Every deployment problem follows from one property: you don't know how long a run takes, because you didn't write the path. Why the run must be a job, where state belongs, and why a transparent retry can send the email twice.
Human-in-the-Loop Patterns for AI Agents
Approval is a budget, not a default — a person asked to confirm forty actions reads the first three. Triage by reversibility rather than risk, and why making an action undoable beats gating it every time.
How to Trace and Debug an AI Agent
Print statements don't survive a system whose control flow is decided at runtime by a model. What every span must carry, the OpenTelemetry GenAI conventions, and why most agent failures are tool failures wearing a reasoning costume.
What Changed in MCP 2026-07-28
The handshake is gone, sampling and roots are deprecated, and HTTP+SSE is no longer a transport. A test for whether the MCP guide you're reading describes a protocol that still exists.
MCP Authorization and OAuth
Optional until you opt in, then very specific. Why the resource parameter is mandatory even when the server ignores it, why PKCE doesn't stop mix-up attacks, and why Dynamic Client Registration just got demoted.
MCP Security
Confused deputy, token passthrough, SSRF through OAuth discovery, and the javascript: URL that becomes remote code execution. The specification names these attacks and states the mitigations — most are not prompt injection.
MCP Transports: stdio vs Streamable HTTP
There are two transports now, not three, and the one that vanished is the one most tutorials still document. What each is exposed to, why they fail in opposite directions, and how statelessness changed the operational story.
How to Connect an MCP Server
Every guide teaches you to build a server; almost none explain connecting one. The config is the same object everywhere — clients just disagree about the wrapper key, which is why your last paste silently did nothing.
MCP Tools, Resources and Prompts
Three server primitives, divided not by what they do but by who pulls the trigger: the model picks tools, the application supplies resources, the user invokes prompts. Choose by blast radius, not by data shape.
Gemini vs ChatGPT
On context, multimodality and tool use they've converged. What hasn't: Gemini Flash's promotional rate doubles on 1 January 2027, Gemini Pro charges double above 200k tokens, and GPT-5.6 Luna undercuts everything at $0.20 per million.