20 practical guides
AI Agents and Automation Guides
Learn how to choose an agent workflow, connect tools and knowledge, keep people in control, and evaluate the complete system.
For operators, developers, and professionals automating multi-step work.
Agents vs chatbots
Choose a conversational assistant or an action-taking agent by comparing task boundaries, tool access, risk, and oversight needs.
How tool calling works
Understand tool schemas, model requests, application execution, returned observations, and the controls around every AI tool call.
Reliable agent workflows
Turn a broad automation goal into bounded states, explicit tools, completion checks, recovery paths, and observable handoffs.
RAG from first principles
Build a clear mental model of indexing, retrieval, grounded generation, citations, and the separate checks each RAG stage needs.
Embeddings and search
Learn how vectors support meaning-based retrieval, where similarity fails, and how to evaluate search against real information needs.
Choosing a vector database
Compare vector storage options using retrieval needs, filters, operations, security, migration, and measured workload evidence.
Chunking for retrieval
Divide documents into retrievable units that preserve meaning, provenance, permissions, and enough context to answer real questions.
Agent memory
Design short-term state and durable memory with purpose limits, provenance, user controls, retrieval rules, and deletion paths.
Human approval for agents
Place informed approval before consequential AI actions with clear previews, authority checks, expiration, and audit evidence.
Source-tracked research agent
Design research automation that decomposes questions, records claim-level evidence, exposes gaps, and produces inspectable citations.
Email triage agent
Classify incoming email, extract actionable facts, route uncertain cases, and keep sending or destructive actions under explicit control.
Meeting follow-up workflow
Convert transcripts into reviewable decisions and actions while preserving uncertainty, consent, source locations, and owner approval.
Spreadsheet analysis agent
Analyze workbook data with explicit scope, deterministic calculations, cell-level evidence, validation, and controlled write-back.
Customer support agent
Ground support answers in approved policy and account data, constrain actions, route uncertainty, and evaluate real resolution quality.
Browser automation agent
Automate browser tasks with restricted sites, observable state, protected credentials, confirmation before side effects, and robust tests.
MCP servers with assistants
Connect assistants to Model Context Protocol servers by defining trust boundaries, permissions, discovery, approval, and operational checks.
Prompt injection defense
Reduce prompt injection risk with explicit trust boundaries, least-privilege tools, validated actions, isolated content, and adversarial tests.
Repeatable agent evaluations
Build repeatable agent evaluations that score decisions, tool use, safety, state changes, and outcomes across representative cases.
Monitoring agent failures
Observe agent outcomes, trajectories, costs, policy events, and recovery signals without collecting excessive sensitive data.
Production agent architecture
Design production agent boundaries for orchestration, tools, state, knowledge, policy, evaluation, observability, and recovery.