Model selection
Choose platform or BYOK credentials, a function-calling-capable model, reasoning effort, and fast mode for an Action Agent without assuming automatic fallback.
Model selection
The Action Agent card in Action Agent Studio stores the system prompt, one model selection, and the reasoning/fast settings. The Agent runtime always constructs a model tool set when capabilities are attached, so a model used for Knowledge, named HTTP, builtin HTTP, or MCP must support function calling.
There is no public Action Agents API or SDK method that calls a model directly. Save a model on the Agent, then send a message. Flow graphs use separate AI nodes documented under ActionFlow.
Platform and BYOK modes
The Agent stores one of two mutually exclusive model selections:
| Mode | Selection | Model credentials | Chat billing |
|---|---|---|---|
| Platform | An enabled PlatformModel routed through LiteLLM | Platform-managed provider credentials on ActionFlows servers | Uses ActionFlows organization platform chat credits |
| BYOK | An enabled Service plus enabled ServiceModel | The organization's provider credential for that service | Does not use ActionFlows platform chat credits; the provider can bill the organization directly |
Platform and BYOK model fields cannot be populated at the same time. Switching modes clears the other mode's fields.
For BYOK, add the provider key in Organization → Integrations → AI Vendors before testing. ActionFlows keeps the key server-side and does not return it as Agent metadata. Named tools, MCP servers, and remote APIs can incur separate external usage even for a platform-backed chat.
Function-calling requirement
Attaching a tool makes it available to the model; it does not force the model to call it. The selected model/provider must be able to emit and continue tool calls reliably.
Current picker behavior differs by mode:
- BYOK: Action Agent Studio lists only enabled, non-deleted
ServiceModelrecords whosesupportsmetadata includes the exactModelCapability.functionCallingvalue. - Platform: Action Agent Studio lists enabled
PlatformModelrecords with accepted chat model types (or no explicit type).PlatformModelcurrently has nofunctionCallingcapability field, so this picker does not prove tool support from metadata.
For a platform model, verify that the specific model supports tool use before attaching tools. A provider or model that does not support the requested tool protocol can fail the turn even though the Agent configuration is otherwise valid.
There is no automatic model fallback. If the selected model is disabled, missing, credential-free, incompatible with tool calls, or temporarily unavailable, fix the saved selection or credentials explicitly.
Effort and fast mode
Action Agent Studio stores effort as low, medium, or high, and fast as a boolean. These settings affect the Agent runtime only when the selected model's catalog metadata says it supports reasoning.
For reasoning-capable models, the current runtime mapping is:
| Setting | Maximum model steps |
|---|---|
fast enabled | 8 |
low | 10 |
medium | 15 |
high | 20 |
fast takes precedence over the selected effort for the step ceiling. The selected effort is also translated into provider reasoning options where the provider has a supported option bag. ActionFlows does not pass a temperature to streamAgentText for reasoning-capable models.
For a model without reasoning support, ActionFlows ignores both effort and fast, uses 15 maximum steps, and explicitly passes temperature 0.5.
These values tune the Agent runtime; they do not guarantee a provider's own reasoning budget, latency, or cost. Provider/model documentation remains authoritative for those characteristics.
Save and refresh behavior
- Select the platform or BYOK mode.
- Choose a compatible model.
- Set effort or fast mode when supported.
- Save and start a new message.
An Agent-owned model or prompt write increments configVersion. The runtime reloads configuration at multiple phases, so a later phase can observe a saved change during an in-flight turn; a turn is not pinned to an immutable configuration snapshot.
For a BYOK model, Service and ServiceModel metadata is catalog data outside the Agent version. After a catalog edit, allow up to the current 60-second configuration cache TTL for the runtime to converge.
Readiness checklist
Before diagnosing a provider error, confirm:
- exactly one complete platform or BYOK model selection is saved;
- the selected catalog row is enabled and not deleted;
- BYOK has an organization credential for the selected service, or the platform worker has its managed provider key;
- a platform-backed Agent has organization credits;
- a tool-using Agent runs on a function-calling-capable model; and
- the model context can hold the base prompt, attached skills, Knowledge/tool interactions, and conversation history.
See Limits for billing-independent runtime limits and Troubleshooting for safe diagnostics.
Related
Knowledge
Attach organization Knowledge Base collections to an Action Agent, monitor extraction, and use bounded search and document-read tools.
Skills
Attach read-only platform and organization-owned prompt skills to an Action Agent and control their system-prompt injection order and refresh behavior.