nibzard/awesome-agentic-patterns

This is a curated catalogue of design patterns for building AI agents. The focus is on practical, repeatable techniques that show up in real products, not toy demos. Each entry must meet three criteria: more than one team is using it, it improves how an agent perceives, reasons, or acts, and it is backed by a public reference such as a blog post, talk, paper, or repository. The patterns are grouped into eight categories. Context and Memory covers approaches to managing what information an agent holds during a session, such as sliding-window curation, episodic memory retrieval, and prompt caching. Feedback Loops covers techniques where the agent receives signals to self-correct, including CI-based feedback, reflection loops, and self-critique evaluators. Learning and Adaptation covers reinforcement fine-tuning and skill library evolution. Orchestration and Control covers how agents decompose tasks, spawn sub-agents, route between tools, and manage cost budgets. Reliability and Eval covers guardrails, evaluation harnesses, logging, and reproducibility. Security and Safety covers isolated execution environments, PII handling, and security scanning. Tool Use and Environment covers shell, browser, and database integration patterns. UX and Collaboration covers prompt hand-offs, staged commits, and async background agent patterns. Each pattern lives in its own markdown file in the patterns folder of the repository. A companion website at agentic-patterns.com provides additional ways to browse the collection, including a compare tool for putting multiple patterns side by side, a decision explorer to match patterns to use cases, and a graph view of how patterns relate to each other. The site is built with Astro and deployed on Vercel, and its source code is included in the repository.