Practice: Many agents working together: multi-agent systems
Self-check
Section titled “Self-check”Seven short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. What is a multi-agent system, in one sentence?
Show answer
Several agents, each with its own role, instructions, and often its own focused tools, working on parts of a larger task and coordinating. Usually a coordinator delegates pieces to specialists, or a pipeline hands off from one agent to the next. Each agent is still just a model in a loop with tools.
2. The lesson says the real question is “fit, not ranking.” What does that mean?
Show answer
The question is never “are multi-agent systems better than single agents?” It is “does this task split into distinct enough jobs that specialized agents beat one well-described generalist, and is that worth the coordination cost?” Whether to split is a property of the task, not a ranking of the patterns.
3. Name the three advantages of splitting work across several agents.
Show answer
Specialization (a focused agent with a small, well-described toolbox picks more reliably than an overloaded generalist), parallelism (independent sub-tasks run at once), and modularity (build, test, and replace each agent independently).
4. What is the cost the “brochures” leave out, and what are its forms?
Show answer
Coordination. It shows up as communication overhead (context lost or distorted at every handoff), more failure points (N agents plus the connections between them), who-decides-what ambiguity (assigning work, resolving disagreement, knowing when done), and harder debugging plus more latency.
5. What is the most common way a multi-agent system actually fails?
Show answer
A bad handoff, not a bad agent. Two agents that each understand the task perfectly can still fail at the seam between them, where one fails to pass a constraint the next one needed. The system is only as reliable as its shakiest handoff.
6. What is the practical path the lesson recommends for getting the tradeoff right?
Show answer
Do not decide it upfront. Build one generalist agent first and run it. Split into multiple agents only when you hit a concrete wall that splitting solves: the toolbox grew so large the agent picks the wrong tool, sub-tasks need to run in parallel, or two parts of the job need instructions that fight each other inside one agent.
7. Someone says, “We moved from one agent to a five-agent system, so it is smarter now.” What is wrong with that?
Show answer
More agents is not more intelligence. A five-agent system is more specialized and more complex than a one-agent system, which are different things from smarter. The individual agents are the same models in loops; you added specialization and coordination cost, not capability.
Try it yourself: specialists or one generalist?
Section titled “Try it yourself: specialists or one generalist?”No tooling, no cost; this is design judgment. For each task, decide whether it splits cleanly enough to warrant specialized agents or is better as one well-designed generalist, and say why in one line. Note: the goal is to fit the design to the task, not to rank the patterns.
A. Answer a single customer question from one help-center article.B. Run a newsroom: research, fact-check, and write fifty briefs a day across different beats, with pieces that can run in parallel.C. Translate a paragraph, then summarize the translation. Two sequential steps, tightly shared context.D. Process a loan application: one agent pulls credit data, one checks fraud signals, one assesses risk, each needing different tools and knowledge.Show answer
- A: one generalist. A small, sequential task; any coordination at all is pure overhead.
- B: specialists. The jobs are genuinely different specialties and the pieces parallelize across beats; specialization and parallelism start to repay the coordination cost.
- C: one generalist. Mostly sequential and the two steps share so much context that handing it between agents would lose more than specialization gains.
- D: specialists. The sub-tasks need different tools, different knowledge, and different instructions; that is the case where splitting into focused agents beats one agent juggling everything.
The deciding question each time is fit: do the jobs split into distinct specialties (and parallelize), or do they share context and stay sequential? Not “which pattern is better.”
Try it yourself: name the coordination cost
Section titled “Try it yourself: name the coordination cost”You are tempted to split a task into a coordinator plus three specialist agents. Before you do, name the coordination cost you are taking on, the way the lesson did for the research-brief team. List at least three specific costs for this split.
Split: coordinator + flights agent + lodging agent + itinerary agent, for planning one short trip.
Show answer
For one short trip, the costs outweigh the gains. Specific costs of this split:
- Communication overhead / lost context at the seams: the budget and date constraints have to pass cleanly from coordinator to each specialist; the itinerary agent failing to receive the chosen dates is a classic seam failure.
- More failure points: four agents plus the handoffs between them, versus one place to go wrong in a single generalist.
- Who-decides-what logic: something must assign each job, resolve conflicts (the lodging agent wants dates the flights agent ruled out), and decide when the trip plan is done.
- Debugging and latency: a bad trip plan means tracing which agent and which handoff; every coordination round adds delay.
For one short trip, a single generalist agent with flight, lodging, and itinerary tools has no seams to lose context across and one place to debug. The split would only start to pay if you were planning many trips in parallel or each specialty genuinely needed knowledge that fights inside one agent.
Flashcards
Section titled “Flashcards”Ten cards. Click any card to reveal the answer. Use the Print flashcards button to lay out the full set as one card per page for offline review.
Q. What is a multi-agent system?
Several specialized agents that coordinate on parts of a larger task, usually via a delegating coordinator or a handoff pipeline. Each agent is still just a model in a loop with tools.
Q. What is the 'fit, not ranking' question?
Not “are multi-agent systems better than single agents?” but “does this task split into distinct enough jobs that specialists beat one generalist, and is that worth the coordination cost?” The task decides, not the prestige of the pattern.
Q. What are the three advantages of multiple agents?
Specialization (a focused agent with a small toolbox picks reliably), parallelism (independent sub-tasks run at once), and modularity (build, test, and replace each agent independently).
Q. What is the coordination cost, in its forms?
Communication overhead (context lost at handoffs), more failure points (N agents plus connections), who-decides-what ambiguity, and harder debugging plus more latency.
Q. When does one generalist agent tend to win?
When the task is mostly sequential, when the pieces share so much context that handoffs lose more than specialization gains, or when the task is small enough that any coordination is pure overhead.
Q. When do multiple agents tend to win?
When the task genuinely splits into different specialties (different tools, knowledge, instructions), when sub-tasks can run in parallel, or when one agent’s combined toolbox is so large it confuses itself.
Q. What is the most common multi-agent failure?
A bad handoff, not a bad agent. The seam between agents, where a constraint one agent needed fails to pass, is where well-built agents still fail.
Q. What practical path gets the tradeoff right?
Build one generalist agent first and run it. Split into multiple agents only when you hit a concrete wall splitting solves (oversized toolbox, a need for parallelism, or instructions that fight inside one agent).
Q. Does adding agents make a system smarter?
No. More agents is more specialization and more complexity, not more intelligence. The individual agents are the same models in loops.
Q. How does multi-agent relate to 'compound systems'?
It is one shape of a compound system (models, retrievers, tools, and agents composed together). The same rule applies to all of them: each component buys capability and costs coordination, so spend the tradeoff on purpose.