Week 4
Many agents
Wednesday, October 21, 11:00–11:50 a.m., Slichter 3853
Joining remotely? The Zoom link is on the home page.
Last week ended with a subagent: one tool call that starts a whole new loop. This week we follow that idea to teams of agents: when to split work, whether your code or the model should decide the steps, how peer agents talk to each other, how to run many agents safely at once, and what the evidence says about when many agents help and when they just cost more.
In class
Four parts. Each card opens the slides at the start of that part; the reading page has the details.
- Part 1SubagentsWhy hand work to another agent, how a subagent is defined, fanning out, and what every hand-off loses.
- Part 2WorkflowsWho decides the next step, your code or the model? Five patterns, and a literature review built from them.
- Part 3Peer agentsTeam shapes, how agents talk (files, messages, a shared server, A2A), what the evidence says, and how teams fail.
- Part 4SandboxesWorktrees, containers and microVMs, keys kept outside the box, and how we tested the week 2 harnesses.
Further reading
The details behind every slide, with sources: subagent and team features in Claude Code and Codex, the multi-agent studies, sandbox options, and A2A.
Homework 4
Give your agent an inbox, and subagents: an email MCP server, packaged as a plugin; a hello to the class agent in a fixed format; and, added October 23, a subagent system in your harness that hands work over, runs subagents in the background, and stops them. Due before class on October 28.