Agentic design patterns primer
A set of practitioner threads lays out concrete agentic design patterns — prompt chaining, routing, reflection, tool use and multi‑agent coordination — as repeatable building blocks. The same authors also share runnable code and an ops‑level five‑layer system for intake‑to‑feedback automation in operations‑heavy businesses. (x.com) (x.com) (x.com)
Agentic artificial intelligence is being packaged into a simple builder’s playbook: break work into steps, send each task to the right path, let the system check itself, call tools, and split jobs across specialists. (vinija.ai) Vinija Jain’s primer lays out those patterns as reusable parts rather than one-off prompts, with sections for prompt chaining, routing, parallelization, reflection, tool use, planning, prioritization, and multi-agent systems. The page also sketches failure modes for each pattern, including bad routing decisions, weak tool contracts, and reflection loops that do not converge. (vinija.ai) In plain terms, prompt chaining is a relay race, where one model output feeds the next step; routing is a traffic cop that sends different inputs to different handlers; reflection is a second pass that critiques or revises an answer before it ships. LangChain’s workflow guide draws the same line between fixed workflows and agents that decide their own next step at runtime. (vinija.ai) (docs.langchain.com) Tool use is the pattern that lets a model reach outside its own text window to search, query a database, or trigger software, and LangChain’s agent docs describe that as a loop of model call, tool call, observation, and stop condition. Multi-agent setups add another layer by assigning separate roles or context windows to separate agents instead of stuffing every instruction into one prompt. (docs.langchain.com) (blog.langchain.com) That framing has spread as teams move from chatbot demos to production systems that need retries, persistence, debugging, and human review. LangGraph’s documentation pitches graph-based execution for exactly that mix of deterministic steps and agent decisions, while LangChain says its higher-level agents are built on top of LangGraph for persistence, streaming, and human-in-the-loop controls. (docs.langchain.com 1) (docs.langchain.com 2) (blog.langchain.com) The operational version of the same idea is a layered system, not a single prompt. A recent architecture write-up from VoltusWave describes five layers for high-volume agent systems: an application gateway, an orchestration layer, a stateless agent pool, a tool and integration layer, and a system of record that holds the state. (voltuswave.ai) That intake-to-feedback pattern is showing up in workflow examples outside software engineering. GitHub projects for onboarding, service intake, and feedback routing all follow the same sequence: capture an inbound request, normalize the data, classify or enrich it with a model, route work to the right team or tool, and track status for review. (github.com 1) (github.com 2) (github.com 3) The catch is that more structure does not remove uncertainty. LangChain wrote on February 25, 2026 that production agents have “infinite” input variety compared with traditional software, which makes traces, evaluations, and monitoring part of the product rather than cleanup work after launch. (blog.langchain.com) The practical message in these threads and docs is narrower than the hype: most teams should start with one agent and good tools, then add routing, reflection, or subagents only when the workload forces it. That turns “agentic” from a slogan into a checklist a builder can test, debug, and run. (blog.langchain.com) (vinija.ai)
