NOTES / AGENTIC DESIGN PATTERNS

Workflows vs agents

The split between fixed LLM pipelines and model-directed agents, and the five workflow patterns that sit between them.

UPDATED 18 NOV 2026#agents#patterns#workflows

The distinction

Anthropic groups everything built from LLM calls under “agentic systems”, then splits that into two kinds1:

  • Workflows: the LLM calls and tools run along code paths you define in advance.
  • Agents: the model decides its own steps and which tools to use, in a loop, until the task is done.

Their main advice is to reach for the simplest thing that works, and they found the most successful teams relied on “simple, composable patterns” rather than heavy frameworks1.

The building block

Both start from an LLM augmented with retrieval, tools and memory1. Everything else is a way of wiring those calls together.

Diagram: an input goes to an LLM, which reads from and writes to retrieval, tools and memory, then produces an output

Diagram: my own, based on the idea described in 1. Filled boxes are LLM calls; dashed boxes are ordinary code.

Five workflow patterns

Pattern Idea Good when
Prompt chaining Break a task into fixed sequential steps, with checks between them The task splits cleanly into stages
Routing Classify the input, then send it to a specialised prompt or model Inputs fall into distinct categories
Parallelisation Run sub-tasks at once (sectioning) or the same task several times (voting) You need speed or more confidence
Orchestrator-workers A central LLM breaks the task down and delegates to workers Sub-tasks can’t be known in advance
Evaluator-optimizer One call generates, another critiques, and they loop There are clear criteria for “better”

All five patterns come from 1. The diagrams below are my own drawings of them.

Prompt chaining

Each step’s output is the next step’s input. A check in code between steps can stop the chain early if something looks wrong.

Diagram: input, LLM step 1, a check that either stops early or passes to LLM steps 2 and 3, then output

Routing

A first call classifies the input. Each category gets its own prompt, tools or even a cheaper or stronger model.

Diagram: input goes to a router, which sends it to one of three handlers, then output

Parallelisation

Either split the work into independent pieces (sectioning), or run the same task several times and compare (voting). Code combines the results.

Diagram: input is split across three parallel LLM calls whose results are aggregated into one output

Orchestrator-workers

Looks like parallelisation, but the orchestrator decides the sub-tasks at run time instead of you fixing them in code.

Diagram: an orchestrator LLM plans and sends work to a variable number of workers, whose results are synthesised into the output

Evaluator-optimizer

One call drafts, another judges against criteria and sends feedback, and the loop repeats until it passes or hits a limit.

Diagram: a generator sends drafts to an evaluator, which returns feedback until it accepts the output

When to use an agent

Agents trade predictability and cost for flexibility. They suit open-ended problems where you can’t hard-code the steps, and they need good tools, clear stopping conditions and guardrails1.

Diagram: a task goes to an LLM that chooses actions in an environment and reads observations back, looping until it is done or hits a limit

Further reading

Gullí’s book catalogues 21 patterns, from prompt chaining and routing to multi-agent collaboration, each with code2.

Sources

  1. [1]Building effective agents — Erik Schluntz and Barry Zhang · Anthropic · 2024PARAPHRASED, SHORT QUOTE
  2. [2]Agentic Design Patterns: A Hands-On Guide to Building Intelligent Systems — Antonio Gullí · Springer · 2025FURTHER READING

© Ashish Kumar. All rights reserved. Quoted and cited material belongs to its owners. How I attribute sources.

Product, company and framework names are trademarks of their respective owners. Mentioning them does not imply affiliation or endorsement.