- Home
- AI agents
- Orchestration and multi-agent systems
- Workflows vs agents
Build · Orchestration and multi-agent systems
AI workflow vs AI agent: which one does your process need?
A workflow runs a process along a path that developers write in advance, calling a model only inside fixed steps. An agent lets the model decide which step comes next and which tools to use. In practice, many enterprise processes are a workflow with one or two agent steps placed where the input varies too much to script.
In this guide: 83 pages
Key takeaways
- Anthropic’s advice is to find the simplest solution possible and add complexity only when needed. Often that means a single model call, or a workflow with model calls inside fixed steps.
- Five workflow patterns cover a lot of ground before you need an agent: prompt chaining, routing, parallelisation, orchestrator-workers and evaluator-optimizer.
- OpenAI’s guide gives an agent the edge when a task involves complex decisions, rules that are hard to maintain or heavy unstructured input. Otherwise, it says, a deterministic solution may suffice.
- Put the agent inside the workflow, with its own tools, a turn limit and an exit, so the flow keeps ownership of approvals, retries and the audit trail.
What is the difference between an AI workflow and an AI agent?
The difference is who decides the next step. In a workflow the developer decides, in code or a visual flow, and the model works inside fixed steps; in an agent the model decides, choosing tools and their order as it goes. Anthropic’s definitions, published on 19 Dec 2024, are the ones most teams now use.
Anthropic: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage.” The same post recommends finding the simplest solution possible and adding complexity only when it is needed. OpenAI’s guide to building agents makes the same point from the other side: an agent beats deterministic automation on complex decision-making, rules that are difficult to maintain and heavy reliance on unstructured data, and otherwise a deterministic solution may suffice.
The difference shows up in operations more than in demos. A workflow follows paths you wrote, so you test each one like software and walk an auditor through it step by step; Microsoft describes Copilot Studio agent flows this way: “The same input always produces the same output.” An agent takes different paths on different days, so you test it with an evaluation set of real tasks, cap its turns, tokens and tools, and log every step to explain a decision afterwards. In Anthropic’s research system, agents used about four times the tokens of a chat, and every extra turn an agent takes adds to what a fixed step would use.
Which workflow patterns handle LLM tasks without an agent?
Anthropic names five patterns that keep the developer in charge of the path: prompt chaining, routing, parallelisation, orchestrator-workers and evaluator-optimizer. Each uses a model where it helps and code everywhere else.
The five workflow patterns from Anthropic’s Building effective agents (19 Dec 2024), with an enterprise use for each.
| Pattern | How it works | Enterprise use | Move to an agent when |
|---|---|---|---|
| Prompt chaining | A fixed sequence of model calls, each working on the previous output, with checks between them | Summarise a contract, check the summary against a clause list, then format it | The sequence changes from one document to the next |
| Routing | A model classifies the input and code sends it down the matching path | Sort inbound tickets into billing, access and hardware queues | New routes appear faster than you can maintain them |
| Parallelisation | Independent model calls run at once, and code merges or votes on the results | Check one marketing asset against several rules at the same time | The checks depend on one another |
| Orchestrator-workers | A model splits the task into subtasks at run time and code collects the results | Research a supplier across filings, news and internal records | The subtasks need tools and decisions of their own |
| Evaluator-optimizer | One model drafts, another critiques against criteria, and the loop repeats until the criteria are met or a cap is reached | Improve a customer reply until it meets tone and policy rules | Each round needs new information found with tools |
Anthropic’s post now carries a banner noting that much of the tooling it describes has changed since December 2024. The patterns have not, and most current frameworks implement them directly as graph or workflow nodes.
When does a step need an agent instead of fixed logic?
When the step cannot be written down as rules without the rules becoming the maintenance problem. Ask four questions of each step, and give it an agent only if the answers point that way.
- Can you list the paths? If the step has five known outcomes, use routing. If new cases keep arriving that nobody planned for, an agent may cost less than maintaining the branches.
- Is the input structured? A form with fixed fields needs extraction, one model call. A thread of emails, attachments and a phone note that must be reconciled needs judgment across several sources.
- What does a wrong action cost? If the step writes to a ledger or contacts a customer, keep the write in the workflow behind an approval and let the agent only propose it.
- Can you test it? An agent step needs an evaluation set of real cases with known outcomes before launch. Without one, you cannot tell whether the agent beats the rule it replaces.
When the answers are mixed, we start with the workflow and promote one step to an agent once the evaluation set shows the rules failing. Moving the other way is normal too: when the agent’s choices settle into a pattern clear enough to write down, that part goes back into code.
How do the platforms combine workflows and agents?
Every major platform now runs both in one product, so the design question is where the boundary sits. The table shows how each lets a deterministic flow call an agent, as documented on 2 Nov 2026.
Deterministic and agent parts by platform, with statuses as the vendor pages showed on 2 Nov 2026.
| Platform | Deterministic part | Agent part | Status |
|---|---|---|---|
| Microsoft Copilot Studio | Agent flows on the standard harness, which Microsoft describes as deterministic | GitHub Copilot harness takes a goal, breaks it into steps and adjusts when a step fails; Customer Conversation harness uses specialised sub-agents | Harnesses page dated 19 Oct 2026; GitHub Copilot harness experience in production-ready preview since Jun 2026 |
| Microsoft Foundry and Agent Framework | Graph workflows with checkpointing; Foundry multi-agent workflows with Sequential, Human-in-the-loop and Group chat templates | Prompt and hosted agents as workflow nodes | Agent Framework 1.0 GA 3 Apr 2026; Foundry multi-agent workflows public preview since 25 Nov 2025 |
| AWS | Step Functions Standard workflows, up to one year with exactly-once semantics, or Lambda durable functions | The AgentCore harness, called from Step Functions through an InvokeHarness state | AgentCore harness GA 17 Jun 2026; Lambda durable functions launched 2 Dec 2025 |
| ADK sequential, parallel and loop workflow agents; ADK 2.0 graph workflows | LLM-driven transfer between agents and sub-agents | ADK Python 2.0 GA 19 May 2026 | |
| Open source | LangGraph graphs; Strands Graph and Workflow; LlamaIndex Workflows | Agents as graph nodes; Strands Swarm for autonomous coordination | LangGraph 1.0 on 22 Oct 2025; Strands Agents 1.0 on 15 Jul 2025 |
Strands’ documentation puts the choice in one line: “Use Graph when you want to define the flow, Swarm when agents should coordinate autonomously, and Workflow when the process is fixed.” RPA vendors moved toward the same design from the other side; UiPath’s platform, launched on 30 Apr 2025, puts agents, robots and people on one system with UiPath Maestro as the orchestration layer.
How do Microsoft, AWS and Google implement it?
Where the deterministic and agent parts live on each cloud, with statuses as the vendor pages showed on 2 Nov 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Deterministic flows | Copilot Studio agent flows; Agent Framework graph workflows | Step Functions; Lambda durable functions | ADK sequential, parallel and loop agents; ADK 2.0 graph workflows |
| Agent inside a flow | Foundry agents as nodes in multi-agent workflows (preview) | AgentCore harness through an InvokeHarness state in Step Functions | LLM agents as sub-agents in an ADK workflow |
| Approval step | Request information action in agent flows (preview from Nov 2025); approval-required tools in Agent Framework | The .waitForTaskToken callback in Step Functions; callbacks in Lambda durable functions | Human input nodes in ADK 2.0 graph workflows |
Copilot Studio’s request information action pauses an agent flow for reviewers through Outlook; it was listed as preview in Nov 2025, so check the current label before relying on it.
Enterprise example · Worked example · illustrative
Invoice exceptions: a workflow with one agent step
Illustrative: a 4,000-person distributor’s accounts payable team handling about 30,000 supplier invoices a month. The numbers are invented.
The workflow keeps doing what it did before: capture each invoice, match it against the purchase order and the goods receipt, and post it. About one invoice in ten fails the match, and today a clerk reads the email thread, the PDF and the receiving note to find out why. That step becomes an agent. It may read the thread, the order, the receipt and the supplier record, and it returns one of four outcomes with a reason and the evidence it used.
Everything around it stays in code. The workflow posts only matched invoices, sends every agent proposal above a set amount to an approver, retries the agent once on a timeout and then routes the case to a clerk. The agent never posts, pays or writes to the supplier.
- Owner
- The accounts payable lead, with a named finance systems engineer for the agent step
- Identity
- Its own agent identity with read access to the ERP, one mailbox folder and the supplier master, and no write rights
- Risk tier
- Medium: it proposes financial outcomes; a person approves those above a threshold and samples the rest
- Approval
- Proposals above the threshold go to an approver inside the workflow; a daily sample of the rest is reviewed
- Evaluation
- 300 past exceptions with known outcomes; the agent must match the clerk on 90 percent before launch
- Retirement
- Reviewed every quarter, and retired if the exception volume no longer justifies it
Frequently asked questions
Is an agentic workflow the same as an agent?
The term is used both ways. Most vendors mean a workflow that contains one or more agent steps, the design this page recommends, while a few use it for any multi-step agent. Ask who decides the next step: if code does, it is a workflow.
Can RPA bots and AI agents work in the same process?
Yes. UiPath’s platform, launched on 30 Apr 2025, puts agents, robots and people on one system with Maestro coordinating them, and Copilot Studio offers computer use next to deterministic agent flows. Keep bots on stable screens and rules, and give an agent the step that needs judgment.
Are agents cheaper to build than workflows?
Sometimes faster to prototype, rarely cheaper to run. An agent spends tokens deciding what a workflow already knows, needs an evaluation set to show it works, and needs monitoring for the paths nobody foresaw. Use one where those costs buy something the rules cannot do.
Should a multi-agent system replace a workflow?
Rarely. Anthropic reports that its multi-agent research system used about 15 times the tokens of a chat, and Cognition argues for a single-threaded agent when steps share context. A workflow that calls one well-scoped agent covers most business processes; the multi-agent patterns page covers the exceptions.
More in Orchestration and multi-agent systems
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- Building effective agents
- A practical guide to building agents
- How we built our multi-agent research system
- Don’t Build Multi-Agents
- Harnesses in Copilot Studio
- Agent flows overview
- What’s new in Copilot Studio
- Introducing multi-agent workflows in Foundry Agent Service
- Microsoft Agent Framework version 1.0
- AgentCore harness
- Top announcements of the AWS Summit in New York 2026
- Choosing workflow type in Step Functions
- Build multi-step applications and AI workflows with AWS Lambda durable functions
- Agent Development Kit: easy to build multi-agent applications
- Agent Development Kit
- Strands multi-agent patterns
- LangChain and LangGraph 1.0
- UiPath launches enterprise platform for agentic automation
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner