AgentDefinition

AgentDefinition<ToolProviderContext, Tools, TOutput> = object

Defined in: packages/core/src/agent/types.ts:72

Type Parameters

ToolProviderContext

ToolProviderContext = unknown

Tools

Tools extends ToolSet = ToolSet

TOutput

TOutput = string

Properties

id

id: string

Defined in: packages/core/src/agent/types.ts:77

---

maxRetries?

optional maxRetries?: number

Defined in: packages/core/src/agent/types.ts:113

AI SDK streamText maxRetries: retryable errors from the request itself, before response streaming starts. Overridable per call via AgentRunInput.maxRetries. Omit to keep the SDK default (2). 0 disables those retries.

---

memory?

optional memory?: AgentMemoryConfig

Defined in: packages/core/src/agent/types.ts:131

---

model?

optional model?: LanguageModel

Defined in: packages/core/src/agent/types.ts:80

Required unless AdlRuntimeConfig.defaults.model is set.

---

outputSchema?

optional outputSchema?: z.ZodType<TOutput>

Defined in: packages/core/src/agent/types.ts:99

Default Zod schema for structured output on every episode. When set, AgentRunResult.output is inferred from the schema; when omitted, TOutput is string and output is the episode text.

---

stopWhen?

optional stopWhen?: AgentStopWhen

Defined in: packages/core/src/agent/types.ts:106

AI SDK stopWhen for this agent’s run / stream. Overridable per call via AgentRunInput.stopWhen. Defaults to DEFAULT_AGENT_STOP_WHEN (stepCountIs(20)). Pass stepCountIs(1) when a workflow should own each model step.

---

streamRetries?

optional streamRetries?: number

Defined in: packages/core/src/agent/types.ts:130

Retry the current model step when the provider stream fails after it started. Overridable per call via AgentRunInput.streamRetries.

Omit to leave this off. 0 disables it. A positive value is forwarded to AI SDK streamText streamRetries (reruns the step on { type: "error" } chunks; earlier completed steps stay). A stream that instead rejects with a retryable HTTP 200 APICallError — a socket close while reading a successful response — is turned into such a chunk, so one budget covers both.

Text is not buffered: a dropped attempt’s partial text may already be on textStream / fullStream / agent_text_delta. The SDK excludes it from the result, structured output, and stored messages, and holds tool calls until an attempt finishes. Requires a model object; a string model id fails with INVALID_INPUT. The last failure still rejects the episode.

---

systemPrompt

systemPrompt: AgentSystemPrompt

Defined in: packages/core/src/agent/types.ts:78

---

titleWorkflow?

optional titleWorkflow?: Workflow<ConversationTitleInput, ConversationTitleOutput>

Defined in: packages/core/src/agent/types.ts:142

Optional workflow that names the conversation after the first successful episode on a new memoryScope. It receives the transcript and must return { title: string }. Pin those types with adl.createWorkflow<ConversationTitleInput, ConversationTitleOutput> (Zod input / output are optional). Failures are ignored.

The runtime runs this workflow with { isolated: true } so it is a separate persisted run and is not nested inside another workflow’s tree. Omit it from adl.config workflows if it should not appear in the inspection UI.

---

tools?

optional tools?: Tools | ToolProvider<Tools, ToolProviderContext>

Defined in: packages/core/src/agent/types.ts:93

A fixed tool set, or a provider resolved once per call — typed directly against ExtendedToolProviderContext<ToolProviderContext> since a definition (unlike the per-call AgentRunInput.tools override) is never widened into a heterogeneous AnyAgent[] registry, so it’s safe to keep ToolProviderContext fully generic here. Overridable per call via AgentRunInput.tools.

The framework never parses or validates toolProviderContext — a ToolProvider that wants Zod validation/defaults calls .parse() itself inside getTools (see createToolProvider’s doc comment for the recipe). Combine several providers, each with their own context needs, via combineToolProviders.