Agent platform
0-day streakReview

Your progress

Level 1

0 XP100 XP to level 2
0day streak
0best streak
0topics started
0cards reviewed
    Review cards

    Progress is stored only in this browser. No account is required.

    Layer 3 · Go deeper · 11 min

    Go deeper

    Size it, build it, then make the calls an interviewer will push you on.

    Back of the envelope

    Size it up

    Tap through each step of the arithmetic.

    How many model and tool calls a second at peak?

      How many runs are executing or waiting at the peak?

        How many input tokens does a 25-step background run send?

          What does that 25-step run cost, with and without caching?

            Build-up · 6 steps

            Build the high-level design

            One functional requirement at a time, ending with a platform that works in one region at modest load, before it is safe or cheap.

            Step 1 of 6

            Define and version

            The agent registry validates a definition, stores it as an immutable version and moves aliases such as prod and canary. Runs are long, so a run started on v7 stays on v7.

            6 interview moments

            You’re the candidate

            The interviewer pushes on the hard parts. Pick an approach and see how it holds up.

            1. Deep dive 1 of 6

              Interviewer

              “A run can last hours, touch money and wait a day for a supervisor. Where does the agent loop run?”

              Your move. Pick an approach:

            2. Deep dive 2 of 6

              Interviewer

              “The bank connects 12 MCP servers exposing 400 tools. Which tools does the model see on each call?”

              Your move. Pick an approach:

            3. Deep dive 3 of 6

              Interviewer

              “Assume a ticket’s hidden text will fool the model sooner or later. Whose authority should its tool calls carry?”

              Your move. Pick an approach:

            4. Deep dive 4 of 6

              Interviewer

              “The analysis agent writes Python to compare 30,000 transactions, next to other tenants’ code. Where does it run?”

              Your move. Pick an approach:

            5. Deep dive 5 of 6

              Interviewer

              “Every call re-sends the history, so a 25-step run sends 600,000 input tokens. What changes the curve, not just the price?”

              Your move. Pick an approach:

            6. Deep dive 6 of 6

              Interviewer

              “A search returns nothing, and the model keeps retrying it. What should the platform do?”

              Your move. Pick an approach:

            Build-up · 8 steps

            Follow one dispute run through it

            The run pins its version, the orchestrator owns the loop, and every action crosses one of two gateways. Follow the ₹12,499 dispute through it.

            Step 1 of 8

            Start

            The bank’s support console calls POST /v1/runs as the signed-in user, with an idempotency key. The run API checks quotas.