An agent is a system in which a language model decides, step by step, which tool to call next and when the task is finished, running in a loop until it succeeds, fails or hits a limit you set. That is the difference from a workflow, where your code fixes the sequence of model calls in advance and the model only fills in each step (Anthropic, Building effective agents). The reason-act-observe loop behind many agent designs is described in the ReAct paper.
In FDE interviews
One candidate for a Google role at L4 reported, in July 2026, that the 60-minute system-architecture round let them choose between an agent design track and an audio/video systems track; they chose agent design, and the interviewer role-played the CTO of a company that wanted an AI-powered system. Source 1FDE Interview Experience at Google (L4) (Blind)PublisherBlindSource typecandidate report on BlindSource 2Google Forward Deployed Engineer Interview Experience (Blind)PublisherBlindSource typecandidate report on BlindSource 3Is Google FDE interview same as SWE? (Blind)PublisherBlindSource typecandidate report on Blind A commenter in a related Blind thread wrote, in August 2026, that the round can stay high level but that they went down into prompts, state machines, retries and handling hallucinations. Source 1FDE Interview Experience at Google (L4) (Blind)PublisherBlindSource typecandidate report on BlindSource 2Google Forward Deployed Engineer Interview Experience (Blind)PublisherBlindSource typecandidate report on BlindSource 3Is Google FDE interview same as SWE? (Blind)PublisherBlindSource typecandidate report on Blind A Blind poster who collects interview experiences for a prep site relayed, in July 2026, a Google Senior FDE phone screen whose agentic system design discussion covered agent orchestration, safety guardrails, stuck agents and infinite loops, and monitoring and human-in-the-loop. Source 2Google Forward Deployed Engineer Interview Experience (Blind)PublisherBlindSource typecandidate report on Blind
A strong answer starts from a workflow and adds autonomy only where the path cannot be fixed in advance, then says how each limit is enforced:
- a step budget set from the traces of successful runs (
max_steps ≈ 2 × p95(steps)), not guessed; - a cost cap per task;
- loop detection that stops the run when the agent repeats a tool call with the same arguments;
- each tool scoped to the permissions of the user the agent acts for;
- every tool with side effects made safe to retry with an idempotency key;
- a handoff to a person with the trace attached.
Then say how you would know it works: score whole trajectories, the tool calls as well as the final answer, on a golden set. The step-budget rule of thumb is ours.
Related: Model Context Protocol, prompt injection, LLM-as-judge.