Every overloaded agent tempts you with the same fix: add more agents. Sometimes that is exactly right. More often you have just bought yourself coordination overhead, a bigger token bill, and a new class of failure that is much harder to debug. This week you learn the patterns, and the honest test for whether you need them.
python3 -m pip install gradio, 300 MB; it will not work on classroom Wi-Fi).Specialization (a narrow role with a narrow tool set), context isolation (each worker gets a clean window instead of one bloated one), and parallelism (four searches at once instead of four in a row). These are real. They are why Anthropic's own research system is multi-agent.
Attribution. A single-agent run has one trace to read. A five-agent run fails somewhere in the seams, and "whose fault was that?" becomes a research question, unless you paid, up front, for per-agent and per-hand-off logging (Week 8's evaluation discipline).
Capability grows roughly with the number of agents. The number of relationships that can go wrong grows faster. This is the oldest result in organization design, and it did not stop being true when the workers became language models.
Pick a team size.
This is the app from page 03: a rule-based router in front of an orders agent and a course-RAG agent, each with its own context window. Press Play to watch a clean hand-off, then the two ways a split system fails: a fact that lives in the wrong agent’s window, and a “yes” that reaches the wrong agent.
Try this: for each failure, decide whether the fix belongs in the router, in a shared state object, or in merging the two agents back into one. Name the trade-off you accept.
Each agent is a recurring per-run expense, not a one-time build cost. Anthropic reports that their multi-agent research system consumed roughly 15× the tokens of a chat interaction (Anthropic Engineering, 2025; figure as reported, approximate). Parallel search buys breadth, not savings. A second agent must earn that line item.
Two agents that disagree do not announce it. They produce two confident, incompatible outputs and let a human discover the contradiction. Your users cannot see the seam. That is a trust failure before it is a technical one.
"Why did the system do that?" has a clean answer only if every agent's trace carries the Week 2 evidence standard (state before → observation → available actions → selected action → result → state after) plus the hand-off payload. Budget observability before you add the agent.
The first step after your proposal: make it run. Milestone I is a working demo of at least one agent on the task you proposed, not a plan, not slides. Submit on iCollege in your own section; the full brief is in the Milestone I dropbox.
Do not add a second agent to Milestone I without a written justification. "It felt cleaner" is not one. The justification you owe is on page 01: name which of the four reasons you are buying (context isolation, permission isolation, parallel breadth, or ownership), and show the pieces can be verified on their own. A single agent with well-designed tools is the correct default, and it is the baseline you will need to prove a second agent helped.
Grading (5 pts): runs on a real input 2 · tool + stop condition 1 · trace + README reproducible 1 · honest status note 1.