Pick a focused question that fits your time, stack, and interview goal.
How much time do you have?
Show one-drill sessions you can finish now.
25 results across 1 active filter
Page 1 of 2
Uses traces to separate unclear goals, bad tool interfaces, missing state, contradictory results, poor termination, and retry amplification.
Chooses delegation patterns based on who owns the user interaction, context, final answer, permissions, and failure handling.
Separates the current model input from authoritative workflow state and selectively retained information across runs.
Combines restrained autonomy, curated tools, durable state, approval, memory, recovery, MCP trust, evaluation, and operational controls.
Uses workflow traces to distinguish planning failure, stale observations, retry ambiguity, and missing terminal conditions in a looping agent.
Uses plans as bounded, revisable execution aids while preserving evidence, policy, and application-owned state transitions.
Defines success, failure, step, time, token, cost, retry, and repetition limits with useful escalation behavior.
Chooses among routing, sequential, parallel, evaluator-optimizer, and bounded agent loops from the task's dependency structure.
Builds a small, task-relevant, typed toolset with clear semantics rather than exposing an entire internal API surface.
Uses provenance, content isolation, constrained authority, behavioral checks, and deterministic enforcement to contain hostile external instructions.
Pauses durable execution with a clear action preview, verifies approver authority, handles expiry and changed state, and resumes idempotently.
Stores only useful, scoped, attributable memories with extraction, verification, retrieval, expiry, correction, and deletion.
Evaluates verified outcomes, trajectory quality, safety, efficiency, recovery, and important slices in realistic environments.
Keeps credentials in trusted executors and replaces broad reusable secrets with scoped references and short-lived authority.
Constrains source-to-sink combinations with least privilege, egress policy, argument checks, approval, and information-flow awareness.
Checkpoints authoritative state, reconciles ambiguous tool outcomes, preserves approvals, and resumes without repeating completed work.
Establishes server trust, audience-bound authorization, least privilege, capability filtering, approval, and safe handling of untrusted results.
Classifies tools by consequence and applies distinct permissions, validation, confirmation, idempotency, and audit controls.
Builds a resumable tool workflow that validates model proposals, authorizes every action, requires approval for side effects, and remains idempotent under retries.
Runs containment and recovery for an agent that ingested malicious instructions from retrieved content and attempted an unauthorized outbound tool call.
Distinguishes model-directed execution from explicit application control and explains when each is appropriate.
Defines MCP as a client-host-server protocol for discovering and using tools, resources, and prompts without confusing interoperability with safety.
Prefers typed APIs for reliable actions and reserves UI automation for constrained cases with stronger observation, approval, and recovery.
Rejects agent complexity when rules, APIs, search, ordinary automation, or a bounded model call solve the task more safely.