Agentic Workflows
Agentic workflows break a task into steps in which language models plan, call tools, check their own output, and revise, instead of answering in one pass.
An agentic workflow is a multistep process in which large language models (LLMs) and tools work toward a goal by iterating: planning, calling tools, checking intermediate results, and revising, instead of producing an answer in a single pass. Andrew Ng popularized the term in March 2024 with a coding example on the HumanEval benchmark: GPT-3.5 answering zero-shot was correct 48.1% of the time and GPT-4 67.0%, but GPT-3.5 wrapped in an iterative agent loop reached up to 95.1%, a larger gain than the jump from GPT-3.5 to GPT-4.
Agentic workflows sit on a spectrum. At one end, a developer fixes the sequence of steps in code and the model makes decisions only inside each step. At the other, an AI agent decides for itself which tools to call and how many steps a task takes. Anthropic's widely cited guide to building effective agents calls the first kind workflows and the second agents, and recommends the simplest structure that solves the task, because predefined paths are cheaper, faster, and easier to test.
Common Workflow Patterns#
Ng grouped agentic design into four patterns: reflection, tool use, planning, and multi-agent collaboration. Production systems usually combine a few concrete building blocks:
- Prompt chaining: Splits a task into fixed steps, each model call working on the previous output, with code checks between steps, such as outlining a report, validating the outline, and then writing it.
- Routing: Classifies an input and sends it to a specialized prompt, tool set, or model, such as cheap models for simple questions and a reasoning model for hard ones.
- Tool use: Lets the model fetch data or act through APIs, databases, code execution, and vision models, returning structured outputs the next step can parse.
- Parallelization: Runs independent subtasks at once or asks several models the same question and votes on the answers.
- Reflection and evaluator-optimizer loops: One call drafts a result and another critiques it against criteria, repeating until the draft passes, the pattern behind Ng's HumanEval gain.
- Orchestrator-workers: A lead model plans subtasks at runtime and delegates them to workers, the step where workflows shade into AI agent orchestration.
A support pipeline that combines routing with an evaluator loop shows how the patterns compose:
Agentic Workflows vs. Related Terms#
- Traditional pipelines: Data and machine learning pipelines run the same steps every time, with no model judgment about what comes next. An agentic workflow uses model calls to make decisions at its branch points.
- Robotic process automation (RPA): Scripted bots repeat recorded clicks and keystrokes on fixed screens. Agentic workflows handle unstructured inputs such as emails, documents, and images.
- Agentic RAG: Applies the same patterns to retrieval, letting the model decide what to search for, judge the results, and search again before answering.
- Compound AI systems: The broader term for systems built from several models, retrievers, and tools, whether or not any step iterates.
Building Reliable Workflows#
Frameworks such as LangChain's LangGraph, CrewAI, the OpenAI Agents SDK, and Microsoft Agent Framework, the successor to AutoGen and Semantic Kernel, provide the graphs, state, and retries that workflows need, and each workflow runs inside an agent harness that executes its tool calls. Reliability comes from the parts around the model: deterministic checks between steps, AI guardrails on inputs and outputs, human approval before irreversible actions, observability traces of every call, and agent evaluation on a fixed task suite whenever a prompt, model, or tool changes. Errors compound across steps, because a workflow whose five steps each succeed 95% of the time completes all five only about 77% of the time.
Agentic Workflows in Computer Vision#
Vision workflows often put a fast detector in front of a slower, more expensive model so the expensive step runs only when it is needed. Ultralytics Platform Agents builds these workflows on a visual canvas from blocks for images and datasets, YOLO26 or deployed models, vision-language models (VLMs), Gate conditions, dataset collection, Slack, webhooks, and export. Its templates show the pattern: the YOLO → VLM Monitor template asks a VLM to describe a scene only when YOLO detects at least one person, Collect Uncertain Detections copies images with confidence scores in a chosen range into a dataset for labeling, and Capture and Alert saves a matching image and posts to Slack from the same condition, with up to 20 blocks per workflow and a Python button that shows the generated code.










