Multi-agent systems and orchestration
- Agents
- Asked often
- About 1 h 30 min to learn
- 3 vetted resources
What this topic covers
Multi-agent systems split a task across several LLM agents, typically an orchestrator that hands work to subagents or passes the conversation to a specialist. Interviewers ask about them because the trade-offs are real and widely debated: separate context windows let agents explore in parallel, but each subagent loses shared context, makes decisions that can conflict, and multiplies token cost. A strong answer names the common patterns (orchestrator-worker, agents as tools, handoffs), explains why read-heavy research tasks parallelise better than write-heavy ones, lists typical failures such as duplicated work, misread instructions and missing verification, and argues when a single agent with good context management is the better choice.
You’re ready when you can
What interviewers check, from the resources below.
- States the two principles: share full context, and actions carry implicit decisions
- Explains with an example how parallel subagents produce conflicting work
- Describes the single-threaded agent plus history-compression alternative
- Argues one case for and one case against going multi-agent
- Contrasts agents as tools (manager keeps control) with handoffs (specialist takes over)
- Says when code-driven orchestration beats letting the LLM decide
- Sketches an evaluator loop and a parallel fan-out in a few lines
- Names the three failure categories with one example each
- Explains why weak or missing verification lets errors compound
- Proposes one design change for a failure mode you name
What to study, in order
Each link opened and checked by hand; the date says when.
Start here
Don't Build Multi-Agents (opens in a new tab)
Cognition · Walden Yan · 20 min · Free · checked 21 Oct 2026
- Why
- “When would you not use multiple agents?” is the question that separates people who have shipped agents from people who have only read about them, and this is the best-known argument for the cautious answer.
- What to read
- Read it all: “Principles of Context Engineering”, “A Theory of Building Long-running Agents”, “Multi-Agents” and “Applying the Principles”. Then reread Anthropic’s multi-agent research system post (already in the library) and list where the two agree and where they disagree.
- What to skip
- Skim “Toward a More General Theory”; it is speculation about the future rather than something you’d use in an answer.
Free alternative
Agent orchestration (opens in a new tab)
OpenAI Agents SDK docs · 20 min · Free · checked 21 Oct 2026
- Why
- It gives you precise names for the patterns, so you can say “agents as tools” versus “handoffs” and explain which one keeps control with the orchestrator.
- What to read
- Read “Orchestrating via LLM” with its “Core SDK patterns” (agents as tools and handoffs), then “Orchestrating via code” (structured-output routing, chaining, evaluator loops, parallel runs with asyncio.gather).
- What to skip
- Skip “Related guides” and the SDK-specific API details; the patterns matter more than the method names.
Go deeper
Why Do Multi-Agent LLM Systems Fail? (opens in a new tab)
arXiv · Mert Cemri, Melissa Z. Pan, Shuyi Yang et al. · 45 min · Free · checked 21 Oct 2026
- Why
- It gives you evidence, not opinion: a taxonomy of 14 failure modes drawn from more than 1,600 annotated traces across seven frameworks.
- What to read
- Read the abstract, the taxonomy figure and the descriptions of the three categories (system design issues, inter-agent misalignment, task verification). Pick two failure modes you could detect in your own traces.
- What to skip
- Skip the annotation-agreement methodology and the LLM-as-a-judge pipeline details unless you are building a failure classifier.
Companies that test Agents most
Put Multi-agent systems and orchestration in your plan
Tell us your interview date and hours a day; it lands on the right day, in the right order.