Skip to content

CreateFlowBuilderAgentOptions

optional agents?: AgentConfig[];

AgentConfig.agents


optional controlModel?: LanguageModel;

Optional model for the control path (routing, decide, extraction), run at temperature 0 for determinism. Defaults to model (the speaker) when unset. Set this to pin control to a reliable provider independent of the speaker.

AgentConfig.controlModel


optional description?: string;

AgentConfig.description


optional experimental?: object;
Name Type Description

outOfBandControl?

boolean

Flow reply nodes: silo flow-transition control tools + deterministic evaluator. Default ON when the agent declares flows; OFF for answering-only agents. Override explicitly to opt out.

AgentConfig.experimental


optional flows?: Flow[];

AgentConfig.flows


optional globalTools?: Record<string, AnyTool>;

Safe, always-available tools made model-visible in EVERY speaking node turn (the agent “base layer”) — e.g. a returns/FAQ knowledge-base lookup the user might ask for mid-flow. This is an explicit allow-list: NEVER put consequential/mutating tools here (they must stay flow-gated), and they are not exposed during non-speaking collect extraction.

AgentConfig.globalTools


optional guardrails?: Guardrails;

AgentConfig.guardrails


optional handoffs?: string[];

AgentConfig.handoffs


host: FlowBuilderHost;

id: string;

AgentConfig.id


optional instructions?: Instructions;

AgentConfig.instructions


optional knowledge?: AgentKnowledge;

AgentConfig.knowledge


optional limits?: Limits;

AgentConfig.limits


optional memory?: AgentMemory;

AgentConfig.memory


optional model?: LanguageModel;

AgentConfig.model


optional name?: string;

AgentConfig.name


optional policy?: Policy;

Decides allow / ask / deny for every tool call this agent makes. Overrides the runtime policy. This is how a delegated worker becomes genuinely read-only: the restriction is enforced at the gate, not by hoping the model respects its instructions.

AgentConfig.policy


optional refine?: RefinementCapability[];

Pre-turn refinement policies. Default: none.

AgentConfig.refine


optional routes?: Route[];

AgentConfig.routes


optional routing?: RoutingPolicy;

AgentConfig.routing


optional skills?: SkillSource;

Bundled procedural skills (Anthropic Agent Skill model): Level-1 name+description in prompt, body via load_skill, resources via read_skill_resource. Scripts = allowedTools only.

AgentConfig.skills


surfaceInstructions: string;

Host-specific policy composed after the shared playbook.


optional tools?: Record<string, AnyTool>;

Durable, model-callable effect tools (exactly-once on replay). Wrap raw AI SDK tools with wrapAiSdkTool().

AgentConfig.tools


optional validate?: ValidationCapability[];

Post-turn validation policies (grounding/confidence gate). Default: none.

AgentConfig.validate


optional workspace?: AgentWorkspaceConfig;

Portable workspace filesystem (or per-session resolver); auto-registers the durable workspace tool when set. Defaults to read-only. { fs, readOnly: false } lets trusted executor code write while the model gets a read-only traversal surface; add modelWritable: true to expose write/edit/mv/rm to the model as well. Filesystem/mount enforcement remains authoritative.

AgentConfig.workspace