University of Michigan’s Latest AI Programming Course “Applied Agentic Software Engineering”

Lecture 14 Notes on “Subagents” Just Released, Structured Around Four Parts: Why Subagents Are Needed, Their Mechanisms, Key Trade-offs, and Differences from Adjacent Concepts.

Notes: https://github.com/eecs498-aase/materials/tree/main/lectures/L14…

Part 1: Why Do We Need Subagents? Because long conversations inevitably degrade, for four reasons: · Context is limited, content only enters but doesn’t exit · Rejected ideas linger in the field (history bleed), and the model picks them up again; this is the source of “degradation” · Every request resends the entire history, caching saves money but not window space · Single conversation is serial, unable to parallelize scheme comparisons when reading code

Solution: Keep only requests and plans in the main conversation; leave reading for each part to its respective subagent; only flow back results, discard all garbage.

Part 2: Subagents Mechanisms · Agent tools are just regular tools to the model; the prompt (i.e., brief) is the sole source of subagent information · Harness immediately replies “launched” (default to background), opens a second conversation: shorter system prompt, fewer tools (Explore without Edit/Write/Agent, physically read-only), same CLAUDE.md, no main conversation history · Ending requires calling SubagentHandback, otherwise content doesn’t reach · Protections when reporting back: mark as model output not user (no user permissions), overall indentation (forged frame lines are distinguishable); use pure formatting to guard conversation boundaries · Brief is a prompt written by the model for the model: position, acceptance criteria, report shape all need to be spelled out clearly; vague briefs yield vague searches

Part 3: Fresh or Forked (Biggest Trade-off) Risks listed by the harness itself: · Bidirectional loss of details: it only knows the brief, you only see the report · Errors and correct discoveries are equally confident · Most dangerous: takes your assumptions, tends to confirm them and send back; for a second opinion, give it code, not conclusions

Fresh: Starting point “brief only”, suitable for “second opinions, independent reviews”, cost “more expensive” Forked: Starting point “full history + brief”, suitable for “side tasks needing all accumulated knowledge”, cost “cheaper (reuses prompt cache)”

Fresh isn’t the cheap option, it’s the independent option. L10’s Goldfish must be fresh: forked ones will fill document gaps with memory, then claim the document is fine.

Part 4: Distinctions and Engineering Three questions distinguish all forms: how many conversations, who talks to whom, who decides the next step. · Subagents: new conversation per segment, communicates only with parent conversation, parent holds authority · Team: multi-conversations + mutual messaging + shared task list (with file locks), you can connect directly to any teammate · Handoff: single conversation, switch to another agent takeover (swap instructions and tools, preserve history), no going back; exactly opposite to subagent · Swarm = mechanisms × scale (Kimi K2.5: 100 concurrent; OpenAI Navier–Stokes: about 10,000 agents, 130 billion tokens) · Parallel editing relies on worktree isolation: each subagent independently checks out + branches, leaves branches after changes for merging

eecs498-aase/materials

Source: https://github.com/eecs498-aase/materials

EECS 498 AASE Materials

Slides and handouts for EECS 498-016, Applied Agentic Software Engineering, University of Michigan, Fall 2026.

Layout

• lectures/ holds one folder per lecture (L01, L02, …), each with a slides.md, a slides.pdf, and a slides.pptx.
• labs/ holds one folder per lab (Lab00, Lab01, …), same shape.
• demos/ holds the codebases the live demos run on, one folder per demo, named for the lecture it belongs to (L01-notes-cli, L03-transcript-analytics). Each has a README and a reset script: run ./reset and it builds you a disposable working copy, configured and ready for aider. Run the demo yourself; that is why they are here.
• egm/ holds the Elephant-Goldfish Model as five skills you can install (peanuts, elephant, goldfish, egm-implement, mean-review), with a walkthrough of each one and a page on how a skill is put together. Start with egm/README.md.

The slides.md is the Slidev source the deck was built from. It is here so you can read, search, or reuse the content. It expects a theme that lives outside this repo, so it will not build as-is.

git clone --depth 1 is the one you want if you only came for a deck or a demo. Most of this repo’s history is compiled slides.

Decks are posted around the time each lecture or lab is delivered. If a week is missing, it has not happened yet.

Links

• Course site: https://eecs498-aase.github.io
• Questions: Ed, or [email protected]