API reference

Use this page when you already understand the Agent SDK workflow and need the exact public entry point or option to reach for.

Entry points

Export	Purpose
`createAgent(definition)`	Define an agent and its default runtime behaviour
`defineTool(definition)`	Define a validated local, manual, approval-gated, or progress-producing tool
`tool(definition)`	Concise alias for `defineTool()`
`createGatewayAgentClient(options)`	Run model turns through Phaseo Gateway
`createAgentDevtools(options)`	Enable local run capture

Agent methods

Method	Returns	Purpose
`agent.run(options)`	`Promise<AgentRunResult>`	Run until completion, pause, stop, cancellation, or failure
`agent.stream(options)`	`AgentStreamResult`	Start the same loop with replayable stream consumers
`agent.continueRun(options)`	`Promise<AgentRunResult>`	Continue saved or paused state with decisions or human input

Stream consumers

Method	Value
`getResult()`	Completed `AgentRunResult`
`getText()`	Complete final text
`getTextStream()`	Text deltas
`getReasoningStream()`	Reasoning deltas when supplied by the model
`getItemsStream()`	Typed `AgentItem<TOutput>` values
`getToolStream()`	Tool lifecycle events
`getFullStream()`	Complete Agent SDK event stream
`cancel(reason?)`	Cancels the active run

Definition options

AgentDefinition includes:

id
model or models
preset
instructions
tools
maxSteps and stopWhen
modelRetry and toolExecution
temperature, maxOutputTokens, and topP
parseOutput and outputSchema
requireApproval and humanReview

Model, instruction, and sampling fields can use fixed or dynamic values.

Run and continuation options

AgentRunOptions supplies input, the model client, application context, optional overrides, cancellation, lifecycle hooks, DevTools, and an application-owned state accessor. AgentContinueOptions additionally accepts:

a prior run or runId
humanInput
approvals and rejections by exact tool-call ID
externally produced toolOutputs

Result records

AgentRunResult contains:

run: durable run status, context, pause details, and stop reason
steps: persisted model/tool step records
stepResults: normalized step results when available
usage: cumulative tokens and reported cost
output: parsed final output
messages: the complete conversation used by the loop
items: ordered typed messages, reasoning, tool activity, errors, and final output

Error types

AgentGatewayError describes gateway or upstream request failures.
AgentSchemaValidationError describes invalid tool, progress, or final-output data.

Start Here

Build

Observe

Reference

Entry points

Agent methods

Stream consumers

Definition options

Run and continuation options

Result records

Error types

​Entry points

​Agent methods

​Stream consumers

​Definition options

​Run and continuation options

​Result records

​Error types

​Related guides

Entry points

Agent methods

Stream consumers

Definition options

Run and continuation options

Result records

Error types

Related guides