What Are AI Agents?
CoveredCore definitions, the agent loop, and when an agent is the right abstraction.
Open current coverageThis map shows the documentation scope this hub is aiming for: what is already covered well, what is only partially covered today, and which GenAI and agent topics should be added next.
Substantive content already exists in the hub.
Covered indirectly, but still missing dedicated depth.
Identified as an important topic for future expansion.
Learn this hub in four passes: GenAI foundations first, then application engineering, then agent systems, then production operations. That order reduces confusion and makes later agent topics much easier to understand.
This is the concrete next-work plan for the repo: prioritized chapters, what each chapter should deliver, and the main files that need to change.
The planned core documentation chapters have been completed. Further work can now focus on incremental updates, framework refreshes, or new case studies as the ecosystem changes.
The core mental models every developer should understand before building agents or GenAI applications.
Core definitions, the agent loop, and when an agent is the right abstraction.
Open current coverageHow models invoke external tools safely and reliably.
Open current coverageSystem prompts, tool instructions, output shaping, and prompt anti-patterns.
Open current coverageHow to choose models by capability, latency, cost, and reliability, then route workloads between them.
Open current coverageSchema-first generation, validation, typed results, and constrained output patterns.
Open current coverageHow prompts, retrieved context, tool results, and memory should be assembled for each run.
Open current coverageWorking with documents, images, audio, and mixed-modal pipelines.
Open current coverageThe application-layer patterns that turn model calls into useful products and internal systems.
Short-term, long-term, episodic, and working memory strategies.
Open current coverageRetrieval pipelines, corrective retrieval, and agent-driven document interaction.
Open current coverageStandardizing tools, resources, and prompts across applications and agents.
Open current coverageA practical decision framework for choosing the right capability layer for a use case.
Open current coverageApproval checkpoints, escalation patterns, and where to keep humans in the control loop.
Open current coverageStreaming, partial results, confidence cues, and graceful failure handling in user-facing apps.
Open current coverageRBAC, tenant isolation, secrets handling, and access control for enterprise GenAI systems.
Open current coverageThe patterns, reasoning strategies, and orchestration approaches used to build capable agentic workflows.
Chain-of-thought, ReAct, plan-and-execute, reflection, and tree-based reasoning.
Open current coverageSupervisor, peer collaboration, and agent-team architectures.
Open current coverageBattle-tested architectures for routing, delegation, collaboration, and tool use.
Open current coverageFramework and runtime tradeoffs across Python, TypeScript, single-agent, and multi-agent systems.
Open current coverageRetries, fallback chains, idempotency, compensation logic, and bounded autonomy.
Open current coverageWhen agents should act through UIs, and the guardrails needed for those systems.
Open current coverageHow agent behavior improves through datasets, regression tests, reviews, and operational feedback.
Open current coverageThe operational capabilities required to run GenAI and agent systems safely, reliably, and cost-effectively.
Validation, filtering, approval gates, PII handling, and prompt-injection defenses.
Open current coverageTracing, structured logs, operational metrics, and alerting for model-driven workflows.
Open current coverageUnit tests, integration tests, benchmark suites, and LLM-as-judge workflows.
Open current coverageScaling, deployment topology, graceful degradation, and runtime concerns.
Open current coverageCaching, model routing, token budgets, attribution, and spend control.
Open current coverageAuditability, approvals, policy enforcement, and compliance-aware GenAI design.
Open current coverageManaging evaluation sets, prompt revisions, and reproducible rollout changes over time.
Open current coverage