EN
Agent Patterns
Patterns that didnβt break production (with code and trade-offs).
Not sure which pattern you need? Design your agent β
ΠΠ°ΡΠ°Π»ΠΎΠ³ ΠΏΠ°ΡΠ΅ΡΠ½ΡΠ²
- ReAct Agentβ β βMaster the ReAct agent pattern: reasoning-action loops, tool use, and guardrails that prevent common failures in production AI workflows.
- Task Decomposition Agentβ β βDecompose complex goals into manageable subtasks so agents can plan, execute, and verify step by step with better reliability.
- Routing Agentβ β βRoute each request to the best agent or tool using explicit criteria, improving accuracy, speed, and cost efficiency in multi-agent systems.
- Orchestrator Agentβ β βLearn orchestration patterns to delegate subtasks, run executors in parallel, track status, and merge outputs into one reliable final result.
- Supervisor Agentβ β βUse a supervisor agent to validate proposed actions, enforce policies, and block unsafe steps before they impact users or systems.
- Multi-Agent Collaborationβ β βUse collaborating agents to split roles, exchange intermediate results, and cross-check outputs to solve complex tasks with higher quality.
- RAG Agentβ β βBuild a RAG agent that finds relevant documents, cites sources, and reduces hallucinations in answers.
- Memory-Augmented Agentβ β βBuild agents that remember user facts and prior outcomes across sessions to deliver consistent, personalized responses without losing control.
- Reflection Agentβ β βAdd one short review pass to catch obvious mistakes before answering, without endless rewrites.
- Self-Critique Agentβ β βRun a safe self-critique loop: one schema-based review, one constrained revision, and a change log for stable quality.
- Fallback Recovery Agentβ β βBuild an agent that recovers from tool and model failures with fallback strategies, retries, and controlled degradation.
- Guarded Policy Agentβ β βImplement a policy-gate that allows, denies, rewrites, or escalates risky agent actions for safe and auditable execution.
- Code Execution Agentβ β βHow an agent runs code in a sandbox to compute reliably, validate hypotheses, and automate tasks with production guardrails.
- Data Analysis Agentβ β βHow an agent ingests, cleans, analyzes, and validates data to produce reproducible metrics and conclusions for decisions.
- Research Agentβ β βUse a bounded research pipeline: search, read, extract facts, and synthesize with citations without tool spam or infinite loops.
AI Agent Patterns for Production Systems
This hub maps practical agent design choices to real production needs: bounded execution loops, policy checks, failure recovery, and source-grounded answers.
Core pages in this section
- ReAct Agent Pattern: Reliable Task Handling
- Supervisor Agent Pattern: Policy Control and Risk Stops
- RAG Agent Pattern: Source-Grounded Answers
- Orchestrator Agent Pattern: Coordinate Multi-Agent Workflows
- Guarded-Policy Agent Pattern: Safe Actions Under Policy
FAQ
What is the best way to go through this section?
Start with the first baseline article, continue with the core pages below, and then validate the ideas with runnable examples.
Can I jump directly to advanced pages?
Yes, but you'll move faster with fewer gaps if you cover the core pages in this section first.
How does this section connect to production work?
These guides map directly to production concerns: architecture, governance, failure modes, and implementation-ready code patterns.