Agentic LLM design patterns
A primer on agentic systems lays out reusable patterns—prompt chaining, routing, parallelization, reflection, tool use, planning and multi-agent coordination—for building complex LLM-based workflows. The write-up draws examples from Anthropic Claude, Google Gemini and OpenAI ChatGPT to illustrate how patterns compose for iterative tasks. (x.com)
Large language model agents are moving from one big prompt to reusable workflows: chains, routers, tool calls, planners, and teams of models. (vinija.ai) A large language model, or LLM, predicts the next token in a sequence. An agent wraps that model with steps, memory, and tools so it can decide what to do next instead of answering in one pass. (learn.microsoft.com) Vinija Jain’s primer groups those workflows into named patterns: prompt chaining, routing, parallelization, reflection, tool use, planning, prioritization, pattern composition, and multi-agent systems. The guide was live on Vinija’s Notes as of April 2026 and presents them as building blocks rather than one full-stack framework. (vinija.ai) Prompt chaining is the simplest pattern: one model call does step one, and its output feeds step two. Microsoft’s agentic workflow guide describes it as a deterministic workflow, where code controls sequence, branching, and retries instead of leaving control to the model. (learn.microsoft.com) Routing sends a task to the best specialist instead of asking one general model to do everything. Anthropic’s December 19, 2024 guide lists routing alongside prompt chaining and parallelization as a core pattern used in effective agent systems. (anthropic.com) Parallelization runs multiple subtasks at the same time and merges the results later. OpenAI’s cookbook example shows focused agents executing concurrently, then passing their outputs to a downstream meta-agent to cut latency and combine perspectives. (developers.openai.com) Reflection adds a review loop: the model drafts, critiques, and revises its own work or hands that review to another model. Vinija’s primer describes reflection as a structured loop, and Anthropic’s companion cookbook implements an evaluator-optimizer workflow for that same idea. (vinija.ai) (github.com) Tool use means the model can call software instead of guessing facts from training data. OpenAI’s Agents SDK says its built-in agent loop handles tool invocation and continues until the task is complete, while Vinija’s guide uses web search and code execution as examples. (openai.github.io) (vinija.ai) Planning breaks a goal into a checklist before execution starts. Vinija’s primer treats planning as its own pattern, and Microsoft draws a line between these code-led workflows and agent-directed loops, where the model chooses tools and order at runtime. (vinija.ai) (learn.microsoft.com) Multi-agent systems split work across several specialists under some coordination scheme. OpenAI’s May 28, 2025 portfolio example uses Macro, Fundamental, and Quantitative agents under a Portfolio Manager, while Vinija’s guide lays out supervisor, network, and hierarchical topologies. (developers.openai.com) (vinija.ai) The point of these patterns is not that every application needs more autonomy. Anthropic says the most successful teams it worked with were using simple, composable patterns rather than complex frameworks, which is also how Jain’s explainer frames the field. (anthropic.com) (vinija.ai)
