Agentic System Design

An “agent” is an artificial intelligence system that can choose steps, use tools, and hand work to specialists instead of answering in one shot. OpenAI says these systems “plan, call tools, collaborate across specialists, and keep enough state to complete multi-step work.” (developers.openai.com) That design work is starting to show up in hiring prep and mock interviews for senior artificial intelligence and platform jobs. On April 17, 2026, Aakash Gupta published a full mock “AI system design interview” built around a “churn reduction agent,” and framed it as the kind of deep technical round candidates now face. (aakashg.com) In Gupta’s mock, the prompt is not “write a prompt” or “ship a chatbot.” The candidate is asked to define requirements, ask clarifying questions, and sketch an end-to-end system that uses signals across mobile and desktop behavior to reduce churn. (aakashg.com) The technical shift is visible in the tooling companies are documenting for builders. OpenAI’s Agents Software Development Kit says the server, not the model alone, owns orchestration, tool execution, approvals, and state, which is the same set of concerns interviewers typically test in a whiteboard design round. (developers.openai.com) Anthropic drew a similar line in a December 19, 2024 engineering post that is still widely cited in 2026 interview prep. It separates “workflows,” where code defines the path, from “agents,” where the model dynamically decides how to use tools and complete the task. (anthropic.com) That distinction changes what a candidate has to explain at the board. A senior answer now has to cover when to use a fixed workflow, when to let a model choose actions, and what latency, cost, and reliability tradeoffs come with each choice. (anthropic.com) Safety checks are part of the design, not an afterthought. OpenAI’s agent documentation breaks guardrails into input, output, and tool guardrails, and says blocking guardrails can stop a run before the expensive model executes or a tool call creates side effects. (openai.github.io) LangChain’s documentation describes the same problem in plainer operational terms: guardrails can block prompt injection, catch sensitive data, enforce business rules, and validate outputs before an agent causes damage. It also splits those checks into deterministic rules and slower model-based reviews, which is the kind of tradeoff interviewers often probe. (docs.langchain.com) The architecture itself is getting simpler in the official guidance, not more elaborate. Anthropic says the most successful teams it worked with used “simple, composable patterns” rather than heavy frameworks, and warned that extra abstraction can make systems harder to debug. (anthropic.com) That leaves candidates preparing for a design conversation that looks more like distributed systems than prompt writing. The strongest answers now describe who owns the reply, how tools are called, where state lives, when a human approves risky actions, and what the system does when the agent fails. (developers.openai.com)