Field notes
Field notes on AI tooling, written from the deep end as I run into it —
mostly from a 4D developer’s chair. There’s a through-line I didn’t
plan: it opens in March 2025 asking whether anyone had tried MCP, and almost
everything since is the same story from a new angle — the ecosystem quietly
settling on a few boring, shared standards. Spawn a process and speak JSON-RPC
(MCP, ACP, the Copilot SDK, Apple’s fm serve); drop a standard
file at the repo root and let any agent read it (AGENTS.md, Agent
Skills). The loud buzzwords — context engineering, loop engineering —
come and go; the standards are what stick. Newest first.
AI agents don't kill workflow engines — they move the DAG down a layer. Agents become the planners; orchestrators like Airflow, Step Functions, and Temporal become the reliable executors.
Xcode 27 speaks the Agent Client Protocol — so you can register the Copilot CLI directly as an agent, pick the model with an env var, and skip the official plugin entirely.
macOS 27 ships the fm CLI (Apple Foundation Models from the terminal), a Python SDK, and Core AI. The kicker: fm serve is an OpenAI-compatible endpoint, so 4D AIKit can drive an on-device model with no API key and no cloud bill.
After prompt, context, and harness engineering comes loop engineering — designing a system that keeps guiding and correcting an LLM until the task is actually done.
An improvised, battle-tested guide to burning fewer tokens in your IDE assistants — compress terminal output, freeze build scripts, script mass edits, throw away long sessions, slim your CLAUDE.md, clean your MCP servers, and let the AI be terse.
Two design-with-AI updates landed together. Claude Design is a workspace you design inside; Google's DESIGN.md is a file your agents read to stay on-brand. One's a destination, the other's a contract — and both touch generating 4D forms.
AGENTS.md is a README for AI agents: a small markdown file at the repo root that hands an LLM the project's shape so it doesn't burn tokens exploring. Here's one auto-generated for 4D projects — plus the Claude Code import trick.
Embedding search is a fast first pass that's only roughly right. A reranker is the cheap precision fix: a cross-encoder that re-scores the top candidates by reading query and document together. Here's why it works and what's out there.
Most MCP servers give your agent access to something external. Xcode 26.3 flips it: Xcode exposes itself as a server, so an agent in VS Code (or anywhere) can discover, build, test, and preview a real Xcode project through Xcode's own machinery.
Chrome ships on-device AI (Gemini Nano) behind Web-standard APIs — translate, summarize, prompt, all local and private. The asterisk: it leans on Google-internal code that isn't in upstream Chromium, so embedded browsers (CEF) don't get it for free.
The Copilot SDK gives four languages programmatic access to the Copilot CLI. Under the hood it's the same trick as MCP, LSP, and ACP — spawn the CLI, talk JSON-RPC — which means any app can host an agent.
Agent Skills landed in VS Code: folders of instructions and scripts an agent loads on demand. The clever part is a three-level progressive loading scheme, so you can keep dozens installed and pay tokens only for the one that's relevant.
Eight months after wondering whether anyone had tried MCP, here it is running for real: the GitHub MCP server in VS Code, wired to our issues, projects, and test cases — with the config committed to the repo so the whole team gets it on checkout.
Context engineering is the step past prompt-writing: assembling everything the model needs — instructions, data, code, memory, tools — in the right format at the right time. A rebrand, sure, but one that points at where the real leverage is.
At WWDC 2025 Apple opened the on-device model behind Apple Intelligence to developers via the Foundation Models framework. The API is tiny — and I wrapped it in a CLI and an MCP server to see how far it goes.
March 2025: MCP is only a few months old and not yet a standard, but people are already building real things on it. Here's the pitch — and a first sketch of what an MCP server for 4D could look like.