What a custom agent is actually metered on
A Workday custom agent draws down usage on two meters at once: one custom agent invocation per request it handles, priced by the model tier, and a separate charge for every Agent-Ready tool call it makes while answering. Building and changing the agent with Developer Agent runs a third meter, Developer Agent invocations. A single question that needs four tool calls is therefore five metered events, and every prompt typed to Developer Agent while building the agent adds invocations before anyone uses it.
Workday's own definitions make the split explicit. A Custom Agent Invocation is processing in Workday's AI runtime "to interpret an end user's request submitted to a production Extend App and complete a tailored workflow in response", and it can use a Base, Standard or Premium model. Tool calls to Agent-Ready Tools are counted on their own and are designated Base or Advanced in Workday's tool call catalog. A Developer Agent Invocation is any query or command submitted to Developer Agent, in the Developer Site chat or through a plugin in a coding tool.
What each meter costs on the Rate Card
The public Flex Credits Rate Card, last updated October 12, 2026, prices these lines under Workday Extend Professional (the SKU Workday's documentation now calls Extend AI):
| Meter | Flex Credits |
|---|---|
| MeterCustom agent invocation, Base tier model | Flex Credits1 per invocation |
| MeterCustom agent invocation, Standard tier model | Flex Credits2 per invocation |
| MeterCustom agent invocation, Premium tier model | Flex Credits3 per invocation |
| MeterBase tool call (data retrieval and inquiry), used by a custom agent | Flex Credits5 per 100 calls |
| MeterAdvanced tool call (transactional or administrative), used by a custom agent | Flex Credits25 per 100 calls |
| MeterPipedream tool call (action in an external system), used by a custom agent | Flex Credits5 per 100 calls |
| MeterDeveloper Agent invocation | Flex Credits2 per invocation, after an annual entitlement of 50,000 |
The Rate Card says these rates are informational and that the customer's agreement sets the final terms. For the general picture of credits, see Workday Flex Credits explained.
Why one request becomes many calls
Four things multiply the count, and all of them are visible in the agent's design or its traces.
The model works in steps. The agent reads the request, picks a tool, reads the result and decides whether it needs another. A request like "show my open Equipment Requests and cancel the oldest" is at least a read and a write. Workday's own sample skill for creating a Work Event lists a worker lookup tool next to the create tools for the same reason.
Write tools pull in prompt tools. Some Advanced tools have required prompts, and Workday supplies read-only prompt tools that return the valid values. Workday's example is changing a preferred name: create_worker_preferred_name_change needs four prompt tools for hereditary suffixes, salutations, social suffixes and titles. The agent has to call those before it can submit the change.
The model tier is set per agent. Workday's tutorial agent shows the agent definition's Model field set to Auto - Base, and Extend AI subscribers can pick other models. A Premium tier model is 3 Flex Credits per invocation against 1 for Base, for every request.
Nothing in the chat shows the count. The user sees one answer, and the number of tool calls behind it is visible in the run's trace.
How to see what a run really did
Open the agent's app in App Builder and select the Agent Traces tab. Each row is one invocation, with its timestamp and a snippet of what the user asked. Click View Trace Details and open the AviatoTurboChatCIS node, which is where the model call happens:
- Its Inputs show the prompt, the conversation history and the tool definitions the model was given.
- Its Outputs show the model's response and every tool call it made, as structured JSON.
Counting the tool calls across a handful of typical requests gives you a real calls-per-request figure to multiply by expected volume, which is more useful than any estimate. Traces show the latest 100 runs and Workday keeps them for 30 days, so capture the figures while you test. Only users with the Admin, Developer or Specialist role can view them.
Where the usage is counted
On the Developer Site, the Entitlements dashboard has an Organization AI Usage entitlement that shows the AI invocations used and the total available. In the tenant, the Platform Consumption Console shows Flex Credit usage by service for UMSA subscribers. Its usage figures update once a day from 2:00 PM UTC for the previous day, and the calculation can take up to 3 hours, so a burst of testing appears the next day, not as you work.
Non-Production use is mostly not charged. Workday defines a Custom Agent Invocation as a request submitted to a production Extend app, and the Platform Consumption Console tracks non-Production tenants and shows an informational balance, but for most offers Workday applies only Production usage against the organization's Flex Credit balance. Workday also notes that UMSA isn't required in non-Production tenants for self-built agents. The Developer Agent entitlement is different: its 50,000 invocations are an annual allowance, prorated over the subscription period, reset on the subscription anniversary and not rolled over, and every Developer Agent prompt counts toward it wherever you are building.
What happens at the limit
Workday doesn't deactivate offers when you exceed your credit entitlement; usage beyond the embedded entitlement is charged against purchased credits. The Alerts and Budget page sends default alerts at 80%, 90% and 100% of the credit threshold to users with the Management Dashboard: Platform Consumption Console domain, and lets you add alerts per agent against a weekly, monthly or annual budget. To stop an agent, deactivate it in Agent Management Hub, which hides it from users and keeps it for reactivation.
A different limit can look like a usage problem: rate limiting. Workday enforces a concurrent request limit of 2,000 on Agent-Ready tool calls, with a refresh rate of 500 and a burst limit of 500. Those are service protections, separate from the credit meters.
Bring the count down
Workday doesn't document these as usage controls; they are recommended design choices that follow from reading the traces. The wider set is in how to reduce Flex Credit consumption, and Flex Credits budget planning covers forecasting.
- Remove tools from the skill's
allowed-toolslist that the traces show the agent never calls; the tool list is what the model receives as input. - Name the tool to use at each step of the skill's instructions, so the model has less to work out.
- Select Read-Only on Custom Agent Actions that only read data; Workday agents ask the user to approve every call to an action that isn't Read-Only, so a lookup that stops for approval adds a turn.
- Test the Base tier before choosing Standard or Premium.
- Do design iteration on Developer Agent in larger, specific prompts rather than many small corrections, since each prompt is an invocation.
The check that a change worked is the same trace comparison: the same request, before and after, with fewer tool calls in the AviatoTurboChatCIS output.