The Desi Architect
Notes / Design Interview Problems / Design an AI Agent Platform (Workflows, Tools, Durable State)
Track 5 · Design Interview Problems

Design an AI Agent Platform (Workflows, Tools, Durable State)

How an enterprise platform runs multi-step LLM reasoning loops, isolates untrusted tool code in microVMs, survives worker crashes without duplicating side effects, and pauses for human sign-off.

How one enterprise platform executes 50 lakh agent workflows daily across 1,000 engineering teams: managing ReAct loops with step budgets, checkpointing state to an event log, sandboxing tools with strict egress filtering, and pausing safely for human approvals.

Interview problem42 min read
On this page
  1. Why this problem is hard
  2. What are we building?
  3. Entities, API and numbers
  4. Building it up, one step at a time
  5. Where it gets hard
  6. Failure modes at a glance
  7. The whole picture
  8. How interviewers judge this
  9. Questions that come next
  10. Saying it out loud in 45 minutes
  11. Sources and assumptions