Posts

Browser MCP snapshot token cost is not the price of one accessibility tree. In practice, it is tool schema tokens, page snapshots, chat history, model output, retries, and browser runtime added together. The right budget number is dollars per completed task, not dollars per million tokens. What are browser MCP snapshots and why do they cost tokens? Browser MCP servers give an LLM a controlled way to inspect and operate a browser. Microsoft’s Playwright MCP is the clearest example: it lets a model interact with pages through structured accessibility snapshots instead of relying only on screenshots or a vision model. That is useful because the model can see buttons, links, roles, labels, and text in a machine-readable form. ...

The best Claude Code dev team stack in 2026 is small: project CLAUDE.md rules, 6-8 focused skills, 4-5 read-first MCP servers, deterministic hooks, and subagents for noisy investigation. I’ve found that teams get more value from governing those layers than from installing every popular skill. That matters because Claude Code has stopped being “just a terminal chatbot.” The current extension surface includes skills, hooks, subagents, agent teams, code intelligence plugins, MCP servers, and packageable plugins. The temptation is to install a marketplace worth of tools on day one. In practice, that usually makes the system slower, harder to trust, and harder to debug. ...

Claude Code subagents are the cleanest way to delegate noisy, self-contained coding work without filling your main session with logs, search results, and half-finished reasoning. In 2026, they are also the simplest entry point to parallel AI development: Markdown files, YAML frontmatter, scoped tools, optional worktree isolation. I use subagents when a task has a clear boundary and a noisy execution path. Running a full test suite, researching three unrelated modules, auditing a diff for security issues, or generating migration notes are good fits. Asking a subagent to co-own an ambiguous refactor with you is usually a bad fit. The difference matters because subagents do not share your full conversation by default. They start with their own prompt, their own context window, and a task summary from the main agent. ...

GitHub Copilot SDK reached GA on June 2, 2026, and it is best understood as an embeddable agent runtime, not a chat wrapper. It gives your app Copilot’s planning, tool calling, file edits, streaming, multi-turn sessions, MCP, hooks, and tracing across TypeScript, Python, Go, .NET, Java, and Rust. I have built enough agent integrations to be suspicious of “SDK” announcements that are mostly a thin prompt API with a nicer package name. The GitHub Copilot SDK is different because it exposes the same runtime layer behind Copilot CLI: your application talks to an SDK client, the SDK talks to Copilot CLI server mode over JSON-RPC, and the runtime handles the agent loop. That matters when you need an agent to inspect files, request permissions, stream events, call tools, and stay useful across multiple turns. ...

If you’re running Copilot Enterprise across a few hundred engineers, you’ve probably stared at the usage dashboard and asked yourself: can I trust these numbers? The short answer is yes, with caveats. The longer answer is what this article is about. I’ve been working with Copilot’s metrics pipeline across several mid-to-large enterprise deployments, and I keep running into the same three questions: How accurate are these metrics? What’s actually covered by the data? And how do I build reliable team-level reports when the API doesn’t give me a team endpoint? This guide tackles each of those head-on. ...

If you’re paying for GitHub Copilot Enterprise across a 200-person engineering org, you need to know whether it’s actually moving the needle. The good news: GitHub now exposes surprisingly detailed usage metrics through dashboards, REST APIs, and NDJSON exports. The bad news: the data has sharp edges that will mislead you if you don’t understand how it’s collected and attributed. I’ve spent the last few months working with Copilot’s metrics pipeline across several enterprise deployments, and this guide covers what I’ve found — the five metric categories, how data actually flows from IDE to dashboard, the API endpoints you’ll need, and the attribution gotchas that will trip you up. ...

If you’re still running a traditional SDLC in 2026, you’re leaving a lot on the table. I’ve been running AI-augmented development workflows for the past year, and the shift from “write every line yourself” to “steer agents that write the lines” is the biggest change in how software gets built since version control went mainstream. The AI-Driven Development Lifecycle (AI-DLC) is the framework that formalizes this shift. It was invented and open-sourced by AWS in 2025, and by mid-2026 it’s become the de facto reference model for teams adopting AI coding agents at scale. This article breaks down what AI-DLC actually is, how the different implementations compare, and what it means for your day-to-day as a developer. ...

If you’re building AI agents that execute code — whether it’s a coding agent running tests, a data analysis pipeline spinning up Python interpreters, or a multi-agent system that needs isolated environments per task — the sandbox you pick directly determines your cost structure, your latency profile, and how much state management code you have to write yourself. Three providers dominate this space in 2026: E2B, Modal, and Fly.io. Each one has a fundamentally different pricing model, cold start story, and persistence strategy. I’ve been running production agent workloads on all three for the past six months, and the differences are bigger than the marketing suggests. ...

I’ve been testing terminal emulators for AI agent workflows since early 2025, and most of them fall into one of two camps: either they’re a traditional terminal with AI features bolted on, or they’re a cloud-dependent IDE pretending to be a terminal. Hyperia, forked from Vercel’s Hyper project and built by a solo developer (kordlessagain / DeepBlue Dynamics), takes a third path that I think is worth paying attention to. ...

What Is an MCP Gateway? And Why You Need One in 2026 The MCP protocol crossed 97 million monthly downloads by mid-2026, and every major AI vendor — Anthropic, OpenAI, Google, Microsoft — ships first-class MCP support. If your team is building AI agents that talk to databases, APIs, or internal tools, you’re already running into the same problem: how do you govern which tools your agents can reach, with whose credentials, and under what audit trail? ...