Workflows vs Cloud Flows vs Agent Flows
Last reviewed: October 15, 2026
Microsoft now gives you three deterministic automation options that look almost identical on the canvas: Power Automate cloud flows, agent flows in Copilot Studio and the new workflows in Copilot Studio. All three have a trigger, actions and connectors. What separates them is where they live, which runtime (harness) runs them and, above all, how they are billed.
TL;DR: Cloud flows are built in Power Automate and paid with Power Automate licenses (Free, Premium per user, Process per flow). Agent flows are built in Copilot Studio on the standard harness and paid with Copilot Credits per action (13 credits per 100 actions), free when a Microsoft 365 Copilot licensed user triggers them through an agent. Workflows are the automation experience of the GitHub Copilot harness in Copilot Studio, with native AI actions, agent handoffs and node-level testing, billed usage-based in Copilot Credits from the moment you start building. No agent involved: cloud flow. A standard agent needs a fixed process as a tool: agent flow. You build on the GitHub Copilot harness: workflow.
What “workflows” in Copilot Studio are
Copilot Studio runs everything on a harness, the runtime between the model and what you build. There are three: the standard harness (rule-based agents and agent flows), the GitHub Copilot harness (reasoning-heavy agents and workflows) and the Copilot chat harness (extending Microsoft Copilot Chat).
Workflows are the automation experience powered by the GitHub Copilot harness. You build them on a redesigned visual canvas with native AI actions, agent handoffs and node-level testing. Like agent flows they are deterministic: the same input follows the same rule-based path.
The GitHub Copilot harness has been generally available since August 3, 2026, and since September 2026 you can add workflows as tools to agents in general availability. Your existing standard agents and agent flows keep running unchanged.
The GitHub Copilot harness is a Copilot Studio framework, not the GitHub Copilot service. Customer data isn’t sent to or processed by GitHub Copilot.
The three at a glance
| Cloud flow | Agent flow | Workflow | |
|---|---|---|---|
| Built in | Power Automate | Copilot Studio (standard harness) | Copilot Studio (GitHub Copilot harness) |
| Paid with | Power Automate license | Copilot Credits per action | Copilot Credits, usage-based |
| Billing starts | With the license | At runtime (tests are free) | When you start building |
| M365 Copilot license covers it | No | Yes, via “When an agent calls the flow” | Yes, inside a standard agent |
| Best for | Classic process automation | Fixed tools for standard agents | Deterministic steps next to reasoning agents |
flowchart TD
A["What are you automating?"] --> B{"Is an agent<br/>involved?"}
B -->|No| C["Cloud flow<br/>Power Automate license"]
B -->|Yes| D{"Which harness<br/>does the agent use?"}
D -->|Standard harness| E["Agent flow<br/>Copilot Credits per action"]
D -->|GitHub Copilot harness| F["Workflow<br/>usage-based Copilot Credits"]
Compared by criterion
What it’s for
- Cloud flow: Automated workflows that connect apps and services, triggered by events or a time of day, with actions across multiple services. The default for process automation without an agent.
- Agent flow: Predefined sequences of flow actions that run repetitive tasks quickly, without agent reasoning at each step. Typically the deterministic tool behind a standard agent, for example a help desk agent that routes simple requests through a fixed process.
- Workflow: Deterministic automation for the GitHub Copilot harness. Microsoft’s reference pattern is accounts payable: the agent reasons through discrepancies and exceptions, while approval, record creation and notification stay in a deterministic workflow.
Where you build it
- Cloud flow: Power Automate (make.powerautomate.com). Create with Copilot from a prompt or from scratch in the new or classic designer.
- Agent flow: Copilot Studio, in the classic designer, either from natural language or by dragging actions onto the canvas. In Copilot Studio, go to Workflows and select New agent flow.
- Workflow: Copilot Studio, on the redesigned canvas. Go to Workflows, select New workflow and add a trigger.
Triggers
- Cloud flow: Automated (an event such as an arriving email), instant (a button) or scheduled.
- Agent flow: Instant, scheduled, other events, or an agent through When an agent calls the flow. Only flows with that trigger can be added to an agent as a tool.
- Workflow: Instant, scheduled, other events or agents. With When an agent calls the flow a workflow becomes a tool in an agent. In the test panel you see the trigger types manual, HTTP request, recurrence, sliding window, polling, connector events and HTTP webhook.
Actions and connectors
- Cloud flow: Standard, premium and custom connectors. Power Automate Free covers standard connectors only. Premium and custom connectors need Power Automate Premium (per user) or a Process license on the flow.
- Agent flow: AI actions (generate text, process documents, run a prompt, call an agent, reply to the calling agent), human in the loop (approvals, requesting information), built-in tools (loops, conditions, data operations, child flows) and connectors including Microsoft 365, third-party and custom connectors. Every executed action counts as an agent flow action.
- Workflow: The same four groups: AI actions including calling an agent, human in the loop (requesting information), built-in tools including child workflows, and connectors including custom connectors.
Billing
- Cloud flow: Power Automate licensing, not Copilot Credits. Cloud flows are not subject to Copilot Studio billing or enforcement.
- Agent flow: 13 Copilot Credits per 100 agent flow actions. Called from a topic, you also pay one classic answer (1 credit). Called through generative orchestration, one agent action (5 credits). Test runs in the designer and the agent’s test chat are free.
- Workflow: Usage-based billing in Copilot Credits for LLM tokens, tools and the harness itself. Building, testing and evaluating already consume credits.
Details and the license matrix follow in Costs and licensing.
Execution and limits
- Cloud flow: A run lasts at most 30 days, including pending approvals. Run history is kept for 30 days. A flow without trigger activity for 90 days can be turned off, unless the owner has a premium license or the flow has a Process license. Daily action limits: 40,000 per Premium user, 250,000 per Process license.
- Agent flow: Runs consume prepaid Copilot Studio capacity. When it is used up, new runs are blocked, running ones complete. Enforcement resets monthly. With pay-as-you-go linked to the environment, there is no block.
- Workflow: Same capacity rule: when prepaid capacity is used up, new workflow runs are blocked, running ones complete. Pay-as-you-go prevents the interruption.
ALM and solutions
- Cloud flow: Can live inside or outside a solution. A Process license and the conversion to an agent flow both require the flow to be in a solution.
- Agent flow: Included in solutions with drafts, versioning, export, import and customization. Version history is stored in Dataverse, with markers for latest, published and past published versions.
Governance
- All three: Data policies (DLP) in the Power Platform admin center classify connectors as Business, Non-business or Blocked. Copilot Studio enforces data policies in real time for all tenants, an exemption is no longer possible.
- Agent flow and workflow: On top, the cost controls for Copilot Studio apply: allocate credits per environment, turn off drawing from the tenant pool, set a monthly limit per agent with Stop usage, and use Azure budgets for pay-as-you-go. These controls apply across Copilot Studio, including workflows.
Error handling and monitoring
- Cloud flow: Run history in Power Automate for 30 days.
- Agent flow: The Flow checker lists all errors, a flow with errors can’t be published. Consumption per flow is in the Power Platform admin center under Licensing > Copilot Studio (Agent flow actions).
- Workflow: The health center banner lists all errors, publishing is blocked until every error is fixed. Test this node runs a single action or AI step in isolation with mocked inputs or values from one of the last 10 runs. The Activity panel is the run history with filters (all, succeeded, failed, running) and shows inputs and outputs per step. To reproduce a production error, open the failed run and replay its upstream data into the failing node.
When to use which
1. SharePoint item, approval, Teams message, no agent involved. Cloud flow. Everything runs on standard connectors with the licenses you already have, no Copilot Credits.
2. HR or help desk agent in Teams that must always create a ticket the same way. Agent flow with the trigger When an agent calls the flow, added as a tool. For employees with a Microsoft 365 Copilot license these runs are free.
3. Agent on the GitHub Copilot harness that reconciles invoices and needs fixed approval and notification steps. Workflow. The agent handles the exceptions, the workflow carries the steps that must run identically every time.
4. Existing cloud flow that should run on your Copilot Studio capacity instead of Power Automate licenses. Convert it to an agent flow. The flow keeps its logic, it consumes Copilot Studio capacity instead of Power Automate billing, and you manage it in Copilot Studio.
5. Nightly sync with thousands of actions per run. Cloud flow with Premium or a Process license. Agent flows bill every action (13 credits per 100), the Power Automate license covers the volume within its daily action limits.
6. New automation in Copilot Studio where single AI steps must be tested and tuned. Workflow. Node-level testing with mocked inputs or replayed run data makes AI steps fast to iterate. Plan for credits from the first build session.
Costs and licensing
| Scenario | What you pay |
|---|---|
| Cloud flow, standard connectors, not shared | Power Automate Free |
| Cloud flow with premium or custom connectors | Power Automate Premium per user, or Process per flow |
| Agent flow called by an agent, user has M365 Copilot | No charge (fair use) |
| Agent flow with any other trigger, or user without M365 Copilot | 13 Copilot Credits per 100 actions |
| Agent flow started by an agent running autonomously | 13 Copilot Credits per 100 actions |
| Workflow inside a standard agent, licensed user | Covered by the M365 Copilot license |
| Workflow in all other cases | Usage-based Copilot Credits, including build and test |
Copilot Credits are charged to the environment where the automation runs, not to the maker or the user. Admins allocate prepaid credits to environments in the Power Platform admin center, or link a pay-as-you-go billing plan to an Azure subscription. Prepaid Copilot Credit packs are shared with Microsoft 365 usage-based experiences, so capacity consumed in Copilot Studio is no longer available there.
Converting a cloud flow to an agent flow
- Open the cloud flow’s detail page in Power Automate.
- Make sure the flow is in a solution. If the Solutions tile is missing, add it to a solution first.
- Make sure the flow sits in the environment that has Copilot Studio capacity (prepaid or pay-as-you-go).
- Select Edit, change the flow’s plan to Copilot Studio, select Save and confirm.
The conversion is one-way, because the billing of the flow changes. You can still open the flow in Power Automate afterwards. A cloud flow can only become an agent flow, not a workflow.
For an agent moving from the standard harness to the GitHub Copilot harness, Microsoft recommends a re-architecture: define the tasks and outcomes, then map topics, flows and prompts to instructions, skills, tools and workflows. A working standard agent doesn’t need to move.
FAQ
Are workflows the same as agent flows?
No. Both are deterministic and look alike, but agent flows run on the standard harness in the classic designer and bill per action. Workflows run on the GitHub Copilot harness in the redesigned designer, add native AI actions, agent handoffs and node-level testing, and bill usage-based from the first build step.
Do cloud flows consume Copilot Credits?
No. Cloud flows use Power Automate licensing and are not subject to Copilot Studio billing or enforcement. They only consume Copilot Credits after you convert them to agent flows.
What happens when the Copilot Credits run out?
New agent flow and workflow runs are blocked, runs already in progress complete. Agents stay available for everything that isn’t a flow. Reallocate capacity, buy more credits or link pay-as-you-go.
Does testing cost money?
Agent flows: no, test runs in the designer and the agent’s test chat are free. Workflows: yes, testing and building consume Copilot Credits.
Can I convert a cloud flow into a workflow?
No. Only the conversion from cloud flow to agent flow exists.
What are workflows in Copilot Studio?
Deterministic automations for the GitHub Copilot harness, built on a redesigned canvas with native AI actions, agent handoffs and node-level testing. They are billed usage-based in Copilot Credits from the moment you start building.
Agent flow vs cloud flow: what’s the difference?
A cloud flow is built in Power Automate and paid with Power Automate licenses. An agent flow is built in Copilot Studio on the standard harness and paid with Copilot Credits per action. You can convert a cloud flow to an agent flow, one-way, once it’s in a solution.
How many Copilot Credits does an agent flow use?
13 Copilot Credits per 100 agent flow actions. Runs through the When an agent calls the flow trigger by a user with a Microsoft 365 Copilot license cost nothing, and test runs in the designer are free.
Related
- Ways to build an agent in Microsoft 365, including the three Copilot Studio harnesses.
- Cloud Flows vs Desktop Flows, plus Premium and Process licenses.
- Standard vs Premium vs Custom connectors
- Flow trigger types
- All Power Automate tips · Knowledge base overview
Sources: Microsoft Learn: Harnesses in Copilot Studio, Workflows overview, Edit and manage your workflow in the designer, Agent flows overview, Edit and manage your agent flow in the designer, Compare Copilot Studio agent harnesses, Billing rates and management, Overview of usage-based billing, Manage costs for agents powered by the GitHub Copilot harness, Configure data policies for agents, Overview of cloud flows, Limits of automated, scheduled, and instant flows, Types of Power Automate licenses. Microsoft Copilot Studio blog: New and improved: GitHub Copilot harness, agent skills and richer context (September 2026). Microsoft Tech Community: More powerful agents and workflows for autonomous business processes (August 3, 2026). Last checked October 15, 2026.