Loading...
AI Agents 1 — all 50 problems
Fifteen implementable exercises drawn from the concrete algorithms behind agent engineering: context trimming, reciprocal rank fusion, memory eviction, tool-schema validation, parallel-call scheduling, circuit breakers, ReWOO plan resolution, tree-of-thoughts beam search, token cost accounting and pass@k scoring.
Agent Loop Control & Planning
- Audit an Agent Trace for Step Limits and LoopsEasy
- Detect a Stuck Agent LoopEasy
- Should the Agent Loop ContinueEasy
- Resolve a ReWOO Plan into Execution WavesMedium
- Resolve a Plan DAG into Parallel WavesMedium
- Reflexion Retry DecisionMedium
- Greedy Subgoal Scheduling Under a Step BudgetMedium
- Detect Repeated State Cycles in a TrajectoryMedium
- Beam Search over a Tree of ThoughtsHard
- Beam Search over Partial PlansHard
Context & Agent Memory
- Trim a Message History to a Token BudgetEasy
- Estimate Token Count of a Message ListEasy
- Sliding Window Keeping the System MessageEasy
- Reciprocal Rank Fusion over Hybrid RetrieversMedium
- Deduplicate and Evict Agent MemoriesMedium
- Trigger Summarization at a Budget FractionMedium
- Recency and Relevance Memory ScoringMedium
- Chunk Text into Overlapping WindowsMedium
- Maximal Marginal Relevance SelectionMedium
- Priority Context Packing Under a Token BudgetHard
Production: Cost, Caching & Evals
- Token Accounting Across a Multi-Model Agent RunEasy
- Dollar Cost of a Token-Priced CallEasy
- Cache Hit RateEasy
- Semantic Cache Hit Rate at a Similarity ThresholdMedium
- Savings from a Semantic CacheMedium
- pass@k EstimatorMedium
- Route to the Cheapest Model Meeting QualityMedium
- Bootstrap Confidence Interval Width for Eval AccuracyMedium
- pass@k Scorecard for an Agent Eval SuiteHard
- Cheapest Cache TTL Meeting a Freshness SLAHard
Reliability & Failure Handling
- Exponential Backoff Schedule with Jitter BoundsEasy
- Classify an Error as RetryableEasy
- Deadline Remaining Before a Tool CallEasy
- Circuit Breaker State Machine for a Failing ToolMedium
- Fallback Chain under a Deadline and Step BudgetMedium
- Total Retry Latency with Capped BackoffMedium
- Circuit Breaker Transitions Over an Event StreamMedium
- Idempotency Deduplication of Retried RequestsMedium
- Timeout Budget Split Across Sequential StepsMedium
- Hedged Requests Expected Completion TimeHard
Tool Use & Function Calling
- Validate a Tool Call Against Its SchemaEasy
- Coerce Tool Arguments to Declared TypesEasy
- Check Required Parameters Are PresentEasy
- Build a Terse, Honest Tool-Result EnvelopeMedium
- Repair Truncated Tool-Call JSONMedium
- Route a Tool Name with Fuzzy MatchingMedium
- Substitute Prior Results into Tool ArgumentsMedium
- Enforce Per-Tool Rate Limits Over a Call StreamMedium
- Schedule Parallel-Safe Tool Calls into WavesHard
- Topologically Order Tool Calls with Argument DependenciesHard