Building AI agents: practical guides

Topics

• Prompt and instruction design

  Prompting and system design for AI agents

  • Writing effective system prompts
  • Designing and describing tools
  • Structuring prompts for cache
  • Injecting dynamic data safely
  • Versioning prompts
• Structured output and schemas

  Structured output and schema validation

  • JSON schema patterns for LLM output
  • Repairing broken LLM JSON
  • Structured output vs function calling
  • Schema compliance across models
  • Generating schemas from code
• Retrieval and agent memory

  Retrieval and memory for AI agents

  • What agent memory should store
  • Embedding and reranker models
  • Chunking and indexing
  • Hybrid search and context budgets
  • Debugging retrieval quality
  • Memory expiry and staleness
• Cost and latency control

  Cost, latency, and budgets for AI agents

  • Estimating and attributing spend
  • Hard caps and refusals
  • Prompt caching and savings
  • Latency budgets and streaming
  • Retry and concurrency budgets
• Agent safety and permissions

  Agent safety, permissions, and reversibility

  • Tool scope and permissions
  • Prompt injection defense
  • Dry runs and idempotency
  • Reversibility and undo
  • Kill switches and denials
• Human review and approvals

  Human in the loop and review workflows

  • Approval flows and prompts
  • Review queues and routing
  • Sampling and reviewer load
  • LLM judges and human baselines
  • Shadow mode and rollout gates
  • Escalation and handoff
• Multi-agent and delegation

  Running multi-agent and delegated work

  • Orchestrator and subagent instructions
  • Delegation budgets and depth
  • Subagent return schemas
  • Tracing multi-agent runs
  • Concurency and deadlines
• Testing and evaluating agents

  Testing and evaluating AI agents

  • Golden tasks and harnesses
  • Fixtures and replay testing
  • Fault injection testing
  • Regression gates
  • Synthetic test cases
  • Tool-call scoring
• Routing and model choice

  Routing and model selection

  • Routing tasks between models
  • Model selection and fallbacks
  • Latency and cost routing
  • Confidence and tie-breaks
  • Routing decision logs
• Tools, integrations and gateways

  Tools, integrations, and gateways

  • Tool contracts and errors
  • Tool gateways
  • Tool errors and retries
  • Streaming and server-sent events