Week 3

The harness around the loop

Wednesday, October 14, 11:00–11:50 a.m., Slichter 3853

Joining remotely? The Zoom link is on the home page.

Last week you built an agent loop with one tool. Codex and Claude Code run the same loop, with layers around it: more tools and MCP servers, instruction files and skills, and the permissions, sandboxes and hooks that decide what the agent may actually do. This week we take each layer in turn and ask one question: does it enter the model's context, when, and at what cost?

In class

Five parts. Each card opens the slides at the start of that part; the reading page has the details.

  1. Part 1Progressive disclosureThe idea of the week: a short index in the context, the library on disk, opened when needed.
  2. Part 2InstructionsWhat fills the context before you type, and AGENTS.md / CLAUDE.md: what to put in it and what it costs.
  3. Part 3Skills and pluginsSkills, the skills CLI, Claude Code plugins and the open Agent Plugins standard, and which to use when.
  4. Part 4Tools, MCP and other protocolsTool design, exactly what is in MCP, and the protocol landscape: MHS for instruments, ACP for editors.
  5. Part 5Control, brieflyPermissions, sandboxes, hooks and the lethal trifecta, with the details in the reading.

Further reading

The details behind every slide, with sources: MCP's 2026 spec, AGENTS.md rules per harness, memory, skills, plugins and channels, permission modes, sandboxes, prompt injection and hooks.

Homework 3

Ask your agent a research question from your phone: two MCP servers, Telegram and one for your own research, packaged as plugins your week 2 harness loads. Due before class on October 21.