Class: Agent

Defined in: agent.ts:982

Agent class - orchestrates LLM interactions with tool use

Example

const agent = new Agent({
  provider: new ClaudeProvider({ apiKey: 'sk-...' }),
  systemPrompt: 'You are a helpful assistant.',
});

agent.registerTool(readFileTool);
agent.registerTool(writeFileTool);

const result = await agent.run('Read the contents of package.json');
console.log(result.response);

Constructors

Constructor

new Agent(config): Agent;

Defined in: agent.ts:1088

Parameters

Parameter	Type
config	AgentConfig

Returns

Agent

Accessors

sessionId

Get Signature

get sessionId(): string;

Defined in: agent.ts:1328

Get the session ID for this agent instance

Returns

string

Methods

addAnchor()

addAnchor(input): Anchor | undefined;

Defined in: agent.ts:1593

Parameters

Parameter	Type
input	AnchorInput

Returns

Anchor | undefined

Deprecated

Use addPin() instead

addGuardrail()

addGuardrail(input): Guardrail | undefined;

Defined in: agent.ts:1648

Add a custom guardrail

Parameters

Parameter	Type	Description
input	GuardrailInput	Guardrail definition

Returns

Guardrail | undefined

The created guardrail, or undefined if guardrails are not enabled

Example

agent.addGuardrail({
  id: 'no-delete-important',
  name: 'Important Files Protection',
  description: 'Prevent deletion of important files',
  patterns: [/rm.*important/i, /delete.*important/i],
  action: 'block',
  message: 'Cannot delete files marked as important',
  scope: ['bash'],
});

addPermission()

addPermission(rule): this;

Defined in: agent.ts:1727

Add a permission rule for a tool

Parameters

Parameter	Type	Description
rule	ToolPermission	Permission rule to add

Returns

this

this for chaining

Example

agent.addPermission({
  toolName: 'bash',
  level: 'once',
  description: 'Execute shell commands',
});

addPin()

addPin(input): Anchor | undefined;

Defined in: agent.ts:1478

Add a pin (critical information that survives context compaction).

Pins are injected into every LLM call and never get compacted. Use them for information that must not be forgotten.

Parameters

Parameter	Type	Description
input	AnchorInput	Pin input (uses AnchorInput type)

Returns

Anchor | undefined

The created pin, or undefined if pins are not enabled

Example

agent.addPin({
  content: 'Team roster: $default, $arch',
  priority: 'info',
  scope: 'session',
});

checkpoint()

checkpoint(metadata?): Promise<string>;

Defined in: agent.ts:2279

Save the current state using the configured checkpointer. Throws if no checkpointer is configured.

Parameters

Parameter	Type	Description
metadata?	Partial<SessionMetadata>	Optional metadata overrides

Returns

Promise<string>

The session ID

Example

const agent = new Agent({
  provider,
  checkpointer: new FileCheckpointer('~/.myapp/sessions/'),
});

await agent.run('Hello!');
const sessionId = await agent.checkpoint();
console.log(`Saved as: ${sessionId}`);

clearAnchors()

clearAnchors(options?): number;

Defined in: agent.ts:1613

Parameters

Parameter	Type
options?	AnchorClearOptions

Returns

number

Deprecated

Use clearPins() instead

clearHistory()

clearHistory(): this;

Defined in: agent.ts:1875

Clear conversation history to start fresh

Returns

this

clearPins()

clearPins(options?): number;

Defined in: agent.ts:1528

Clear pins based on criteria

Parameters

Parameter	Type	Description
options?	AnchorClearOptions	Clear options for filtering which pins to remove

Returns

number

Number of pins removed

clearSessionPermissions()

clearSessionPermissions(): this;

Defined in: agent.ts:1790

Clear all session-level permissions

Returns

this

compact()

compact(options?): Promise<{
  categoryStats?: Record<ContextCategory, {
     action: "preserved" | "compacted" | "summarized";
     tokensAfter: number;
     tokensBefore: number;
  }>;
  filesCreated?: string[];
  messagesPreserved: number;
  originalTokens: number;
  restorationHintsInjected: boolean;
  rounds: number;
  success: boolean;
  summary: string;
  summaryTokens: number;
  toolResultsRepaired: number;
}>;

Defined in: agent.ts:2049

Compact the conversation context to reduce token usage.

This is the recommended way to trigger context compaction externally. It handles:

1. Summarizing older messages
2. Repairing tool use/result pairing (prevents API errors)
3. Injecting context restoration hints (if file tracking is enabled)
4. Updating the conversation history

Parameters

Parameter	Type	Description
options?	{ emergency?: boolean; injectRestorationHints?: boolean; targetUtilization?: number; useSmartCompaction?: boolean; }	Compaction options
options.emergency?	boolean	Use emergency mode (more aggressive summarization). Default: auto-detect based on context utilization
options.injectRestorationHints?	boolean	Inject file restoration hints after compaction. Only applies if file tracking is enabled. Default: true (if file tracking is enabled)
options.targetUtilization?	number	Target utilization after compaction (0-1). Default: from context manager config (typically 0.5)
options.useSmartCompaction?	boolean	Use smart category-aware compaction instead of simple summarization. Smart compaction: - Preserves system and recent messages completely - Saves large tool results to files - Summarizes history with LLM Default: true

Returns

Promise<{ categoryStats?: Record<ContextCategory, { action: "preserved" | "compacted" | "summarized"; tokensAfter: number; tokensBefore: number; }>; filesCreated?: string[]; messagesPreserved: number; originalTokens: number; restorationHintsInjected: boolean; rounds: number; success: boolean; summary: string; summaryTokens: number; toolResultsRepaired: number; }>

Compaction result with statistics

Example

// Basic compaction
const result = await agent.compact();
console.log(`Reduced from ${result.originalTokens} to ${result.summaryTokens} tokens`);

// Compaction without restoration hints
await agent.compact({ injectRestorationHints: false });

// Emergency compaction (more aggressive)
await agent.compact({ emergency: true });

createSubAgent()

createSubAgent(config): this;

Defined in: agent.ts:2463

Create and register a sub-agent with isolated context.

Sub-agents are specialized agents that handle discrete tasks independently. They have their own context window and can have different tools/permissions.

Parameters

Parameter	Type
config	SubAgentConfig

Returns

this

Example

agent.createSubAgent({
  name: 'code-reviewer',
  description: 'Reviews code for security and quality issues',
  systemPrompt: 'You are a code review specialist...',
  tools: [readFileTool], // Restricted tools
  contextMode: 'isolated',
});

const result = await agent.runSubAgent('code-reviewer', 'Review src/auth.ts');

createWithMemory()

static createWithMemory(
   config, 
   memoryOptions?, 
memoryDir?): Promise<Agent>;

Defined in: agent.ts:1307

Create an agent with project memory loaded from files.

This factory method automatically discovers and loads project-specific instructions from files like CLAUDE.md, GEMINI.md, PROJECT.md, etc.

Parameters

Parameter	Type	Description
config	Omit<AgentConfig, "projectMemory">	Agent configuration
memoryOptions?	ProjectMemoryOptions	Project memory loading options
memoryDir?	string	Directory to search for memory files (defaults to cwd)

Returns

Promise<Agent>

Agent instance with loaded project memory

Example

// Load Claude-specific instructions
const agent = await Agent.createWithMemory(
  {
    provider,
    systemPrompt: 'You are a helpful assistant.',
  },
  { providers: 'claude' },
  '/path/to/project'
);

// Load instructions for multiple providers
const agent = await Agent.createWithMemory(
  { provider },
  { providers: ['claude', 'gemini'], includeGeneric: true }
);

// Access loaded memory
const memory = agent.getProjectMemory();
console.log(`Loaded ${memory?.files.length} memory files`);

disableGuardrail()

disableGuardrail(id): boolean;

Defined in: agent.ts:1690

Disable a guardrail by ID

Parameters

Parameter	Type
id	string

Returns

boolean

disposeAllSubAgents()

disposeAllSubAgents(): void;

Defined in: agent.ts:2739

Dispose all sub-agents and release their resources.

Useful for cleanup when the parent agent is done or to free memory during long-running sessions.

Returns

void

disposeSubAgent()

disposeSubAgent(name): boolean;

Defined in: agent.ts:2714

Dispose a sub-agent and release its resources.

This clears the sub-agent’s:

• Conversation history
• Context manager state
• Tool registry

After disposal, the sub-agent must be re-created to use again.

Parameters

Parameter	Type
name	string

Returns

boolean

emitCustomEvent()

emitCustomEvent(config): void;

Defined in: agent.ts:1840

Emit a custom event that will be streamed to event handlers.

This allows tools, middleware, and user code to emit custom events that are streamed alongside built-in agent events.

Inspired by LangGraph’s get_stream_writer() pattern. Addresses issues like LangGraph #6330 (preserve event metadata).

Parameters

Parameter	Type	Description
config	CustomEventConfig	Custom event configuration

Returns

void

Example

agent.emitCustomEvent({
  name: 'progress',
  data: { step: 1, total: 5, message: 'Processing...' },
  metadata: { toolName: 'myTool' },
});

enableGuardrail()

enableGuardrail(id): boolean;

Defined in: agent.ts:1683

Enable a guardrail by ID

Parameters

Parameter	Type
id	string

Returns

boolean

formatRestorationHints()

formatRestorationHints(): string;

Defined in: agent.ts:1978

Format context restoration hints based on tracked file accesses. Returns empty string if no files have been accessed or file tracking is disabled.

Returns

string

fromState()

static fromState(state, options): Agent;

Defined in: agent.ts:2387

Create an agent from a serialized AgentState object.

Parameters

Parameter	Type	Description
state	AgentState	The serialized agent state
options	{ checkpointer?: Checkpointer; onEvent?: AgentEventHandler; provider: LLMProvider; systemPrompt?: string; tools?: Tool<object>[]; }	Options for the new agent
options.checkpointer?	Checkpointer	-
options.onEvent?	AgentEventHandler	-
options.provider	LLMProvider	-
options.systemPrompt?	string	-
options.tools?	Tool<object>[]	-

Returns

Agent

Example

// Load state from somewhere
const json = await fs.readFile('session.json', 'utf-8');
const state = JSON.parse(json);

// Create agent from state
const agent = Agent.fromState(state, { provider });
await agent.run('Continue...');

getAnchor()

getAnchor(id): Anchor | undefined;

Defined in: agent.ts:1597

Parameters

Parameter	Type
id	string

Returns

Anchor | undefined

Deprecated

Use getPin() instead

getAnchorManager()

getAnchorManager(): AnchorManager | undefined;

Defined in: agent.ts:1617

Returns

AnchorManager | undefined

Deprecated

Use getPinManager() instead

getAnchors()

getAnchors(options?): Anchor[];

Defined in: agent.ts:1601

Parameters

Parameter	Type
options?	AnchorQueryOptions

Returns

Anchor[]

Deprecated

Use getPins() instead

getBudgetStatus()

getBudgetStatus(): BudgetStatus | undefined;

Defined in: agent.ts:1415

Get budget status.

Returns

BudgetStatus | undefined

getContextManager()

getContextManager(): ContextManager | undefined;

Defined in: agent.ts:1923

Get the context manager (if configured)

Returns

ContextManager | undefined

getContextStats()

getContextStats(): ContextStats | undefined;

Defined in: agent.ts:1938

Get context statistics

Returns

ContextStats | undefined

getDeadMessagePruneStats()

getDeadMessagePruneStats(): 
  | {
  errorsPruned: number;
  permissionsPruned: number;
  prunedCount: number;
  tokensSaved: number;
}
  | undefined;

Defined in: agent.ts:1954

Get dead message pruning statistics (errors pruned, permissions pruned, tokens saved).

Returns

| { errorsPruned: number; permissionsPruned: number; prunedCount: number; tokensSaved: number; } | undefined

getFileTracker()

getFileTracker(): FileAccessTracker | undefined;

Defined in: agent.ts:1970

Get the file access tracker (if file tracking is enabled)

Returns

FileAccessTracker | undefined

getGuardrail()

getGuardrail(id): Guardrail | undefined;

Defined in: agent.ts:1655

Get a guardrail by ID

Parameters

Parameter	Type
id	string

Returns

Guardrail | undefined

getGuardrailManager()

getGuardrailManager(): GuardrailManager | undefined;

Defined in: agent.ts:1697

Get the guardrail manager (if configured)

Returns

GuardrailManager | undefined

getGuardrails()

getGuardrails(): Guardrail[];

Defined in: agent.ts:1662

Get all guardrails

Returns

Guardrail[]

getHistory()

getHistory(): Message[];

Defined in: agent.ts:1886

Get the current conversation history

Returns

Message[]

getModel()

getModel(): string;

Defined in: agent.ts:1553

Get the current model ID for this agent.

Returns

string

The model ID string (e.g., ‘claude-sonnet-4-20250514’)

getObservationMaskStats()

getObservationMaskStats(): 
  | {
  activeStamps: number;
  inputsCompacted: number;
  maskedCount: number;
  tokensSaved: number;
}
  | undefined;

Defined in: agent.ts:1945

Get observation masking statistics (tokens saved, observations masked, inputs compacted).

Returns

| { activeStamps: number; inputsCompacted: number; maskedCount: number; tokensSaved: number; } | undefined

getPermission()

getPermission(toolName): ToolPermission | undefined;

Defined in: agent.ts:1742

Get a permission rule by tool name

Parameters

Parameter	Type
toolName	string

Returns

ToolPermission | undefined

getPermissionLevel()

getPermissionLevel(toolName): PermissionLevel;

Defined in: agent.ts:1768

Get the effective permission level for a tool

Parameters

Parameter	Type
toolName	string

Returns

PermissionLevel

getPermissionManager()

getPermissionManager(): PermissionManager | undefined;

Defined in: agent.ts:1805

Get the permission manager (if configured)

Returns

PermissionManager | undefined

getPermissions()

getPermissions(): ToolPermission[];

Defined in: agent.ts:1749

Get all permission rules

Returns

ToolPermission[]

getPin()

getPin(id): Anchor | undefined;

Defined in: agent.ts:1489

Get a pin by ID

Parameters

Parameter	Type
id	string

Returns

Anchor | undefined

getPinManager()

getPinManager(): AnchorManager | undefined;

Defined in: agent.ts:1535

Get the pin manager (if configured)

Returns

AnchorManager | undefined

getPins()

getPins(options?): Anchor[];

Defined in: agent.ts:1496

Get all pins, optionally filtered

Parameters

Parameter	Type
options?	AnchorQueryOptions

Returns

Anchor[]

getProjectMemory()

getProjectMemory(): ProjectMemory | undefined;

Defined in: agent.ts:1356

Get the loaded project memory (if any).

Project memory contains instructions loaded from files like CLAUDE.md, GEMINI.md, PROJECT.md, etc.

Returns

ProjectMemory | undefined

The loaded project memory, or undefined if none was loaded

Example

const memory = agent.getProjectMemory();
if (memory) {
  console.log(`Loaded ${memory.files.length} instruction files`);
  console.log(`Total tokens: ~${memory.estimatedTokens}`);
  for (const file of memory.files) {
    console.log(`  - ${file.relativePath}`);
  }
}

getSessionPermissions()

getSessionPermissions(): string[];

Defined in: agent.ts:1798

Get all tools with session-level permission

Returns

string[]

getStreamWriter()

getStreamWriter(eventName?): StreamWriter;

Defined in: agent.ts:1866

Get a stream writer function for emitting custom events.

This returns a simple function that can be passed to tools or middleware for streaming progress updates.

Parameters

Parameter	Type	Default value	Description
eventName	string	'stream'	Name for all events emitted by this writer

Returns

StreamWriter

A stream writer function

Example

const writer = agent.getStreamWriter('myTool');
writer('Starting...', { phase: 'init' });
writer('Processing...', { phase: 'work', progress: 50 });
writer('Done!', { phase: 'complete' });

getSubAgent()

getSubAgent(name): Agent | undefined;

Defined in: agent.ts:2686

Get a registered sub-agent by name

Parameters

Parameter	Type
name	string

Returns

Agent | undefined

getSubAgentNames()

getSubAgentNames(): string[];

Defined in: agent.ts:2693

Get all registered sub-agent names

Returns

string[]

getToolDefinitions()

getToolDefinitions(): ToolDefinition[];

Defined in: agent.ts:2829

Get all registered tool definitions

Returns

ToolDefinition[]

getToolRegistry()

getToolRegistry(): ToolRegistry;

Defined in: agent.ts:1931

Get the tool registry instance. Useful for setting up fallback handlers or inspecting registered tools.

Returns

ToolRegistry

getTotalInputTokens()

getTotalInputTokens(): number;

Defined in: agent.ts:1401

Get total input tokens used across all LLM calls.

Returns

number

getTotalOutputTokens()

getTotalOutputTokens(): number;

Defined in: agent.ts:1408

Get total output tokens used across all LLM calls.

Returns

number

getTotalTokens()

getTotalTokens(): number;

Defined in: agent.ts:1394

Get total tokens used across all LLM calls.

Returns

number

getUsageStats()

getUsageStats(): UsageStats | undefined;

Defined in: agent.ts:1387

Get usage tracking statistics.

Returns

UsageStats | undefined

Usage statistics or undefined if usage tracking is not enabled

Example

const stats = agent.getUsageStats();
if (stats) {
  console.log(`Total calls: ${stats.totalCalls}`);
  console.log(`Total tokens: ${stats.totalTokens}`);
  console.log(`Input tokens: ${stats.totalInputTokens}`);
  console.log(`Output tokens: ${stats.totalOutputTokens}`);
}

getUsageSummary()

getUsageSummary(): string | undefined;

Defined in: agent.ts:1429

Get a human-readable usage summary.

Returns

string | undefined

getVerbosityLevel()

getVerbosityLevel(): VerbosityLevel;

Defined in: agent.ts:1963

Get current verbosity level based on context pressure

Returns

VerbosityLevel

grantSessionPermission()

grantSessionPermission(toolName): this;

Defined in: agent.ts:1775

Grant session-level permission for a tool

Parameters

Parameter	Type
toolName	string

Returns

this

hasAnchor()

hasAnchor(id): boolean;

Defined in: agent.ts:1605

Parameters

Parameter	Type
id	string

Returns

boolean

Deprecated

Use hasPin() instead

hasAnchors()

hasAnchors(): boolean;

Defined in: agent.ts:1621

Returns

boolean

Deprecated

Use hasPins() instead

hasCheckpointer()

hasCheckpointer(): boolean;

Defined in: agent.ts:2292

Check if a checkpointer is configured

Returns

boolean

hasGuardrail()

hasGuardrail(id): boolean;

Defined in: agent.ts:1669

Check if a guardrail exists

Parameters

Parameter	Type
id	string

Returns

boolean

hasGuardrails()

hasGuardrails(): boolean;

Defined in: agent.ts:1704

Check if guardrails are enabled

Returns

boolean

hasPermissions()

hasPermissions(): boolean;

Defined in: agent.ts:1812

Check if permissions are enabled

Returns

boolean

hasPin()

hasPin(id): boolean;

Defined in: agent.ts:1503

Check if a pin exists

Parameters

Parameter	Type
id	string

Returns

boolean

hasPins()

hasPins(): boolean;

Defined in: agent.ts:1542

Check if pins are enabled

Returns

boolean

hasProjectMemory()

hasProjectMemory(): boolean;

Defined in: agent.ts:1363

Check if project memory was loaded

Returns

boolean

hasUsageTracking()

hasUsageTracking(): boolean;

Defined in: agent.ts:1443

Check if usage tracking is enabled.

Returns

boolean

isBudgetExceeded()

isBudgetExceeded(): boolean;

Defined in: agent.ts:1422

Check if budget is exceeded.

Returns

boolean

isToolSilent()

isToolSilent(name): boolean;

Defined in: agent.ts:2836

Check if a tool is marked as silent (no spinner or result output)

Parameters

Parameter	Type
name	string

Returns

boolean

registerTool()

registerTool(tool): this;

Defined in: agent.ts:2808

Register a tool that the agent can call during conversations.

Tools are functions the LLM can invoke to perform actions like reading files, running commands, or querying APIs. The LLM sees the tool’s name, description, and parameter schema, then decides when to call it.

Parameters

Parameter	Type	Description
tool	Tool	Tool definition created with defineTool()

Returns

this

The agent instance (for chaining)

Example

agent.registerTool(defineTool({
  name: 'get_weather',
  description: 'Get current weather for a city',
  parameters: {
    type: 'object',
    properties: { city: { type: 'string' } },
    required: ['city'],
  },
  execute: async ({ city }) => {
    const data = await fetchWeather(city);
    return { content: JSON.stringify(data) };
  },
}));

registerTools()

registerTools(tools): this;

Defined in: agent.ts:2819

Register multiple tools at once.

Parameters

Parameter	Type	Description
tools	Tool<object>[]	Array of tool definitions

Returns

this

The agent instance (for chaining)

removeAnchor()

removeAnchor(id): boolean;

Defined in: agent.ts:1609

Parameters

Parameter	Type
id	string

Returns

boolean

Deprecated

Use removePin() instead

removeGuardrail()

removeGuardrail(id): boolean;

Defined in: agent.ts:1676

Remove a guardrail by ID

Parameters

Parameter	Type
id	string

Returns

boolean

removePermission()

removePermission(toolName): boolean;

Defined in: agent.ts:1735

Remove a permission rule by tool name

Parameters

Parameter	Type
toolName	string

Returns

boolean

removePin()

removePin(id): boolean;

Defined in: agent.ts:1512

Remove a pin by ID

Parameters

Parameter	Type
id	string

Returns

boolean

true if pin was removed, false if not found

removeSubAgent()

removeSubAgent(name): boolean;

Defined in: agent.ts:2700

Remove a registered sub-agent (alias for disposeSubAgent)

Parameters

Parameter	Type
name	string

Returns

boolean

resetUsageTracking()

resetUsageTracking(): void;

Defined in: agent.ts:1436

Reset usage tracking data.

Returns

void

resume()

static resume(sessionId, options): Promise<Agent>;

Defined in: agent.ts:2345

Resume an agent from a saved session.

Parameters

Parameter	Type	Description
sessionId	string	Session ID to resume
options	{ checkpointer: Checkpointer; onEvent?: AgentEventHandler; provider: LLMProvider; systemPrompt?: string; tools?: Tool<object>[]; }	Resume options (provider and checkpointer required)
options.checkpointer	Checkpointer	-
options.onEvent?	AgentEventHandler	-
options.provider	LLMProvider	-
options.systemPrompt?	string	-
options.tools?	Tool<object>[]	-

Returns

Promise<Agent>

Example

const checkpointer = new FileCheckpointer('~/.myapp/sessions/');

// Resume a previous session
const agent = await Agent.resume('session_abc123', {
  provider: new ClaudeProvider({ apiKey: '...' }),
  checkpointer,
});

// Continue the conversation
await agent.run('Continue where we left off...');

revokeSessionPermission()

revokeSessionPermission(toolName): boolean;

Defined in: agent.ts:1783

Revoke session-level permission for a tool

Parameters

Parameter	Type
toolName	string

Returns

boolean

run()

run(userMessage, options?): Promise<AgentRunResult>;

Defined in: agent.ts:2874

Run the agent with a user message and return the result.

This is the main entry point for agent interaction. The agent will:

1. Add the user message to conversation history
2. Send the conversation to the LLM
3. Execute any tool calls the LLM requests
4. Repeat steps 2-3 until the LLM responds with text (no tool calls)
5. Return the final text response and metadata

Events are emitted throughout the process via the onEvent callback configured at construction time.

Parameters

Parameter	Type	Description
userMessage	string | ContentBlock[]	The user’s message (string or content blocks for images)
options?	RunOptions	Optional run configuration (max iterations, abort signal, etc.)

Returns

Promise<AgentRunResult>

The agent’s response, tool call history, and context stats

Examples

const result = await agent.run('What files are in this directory?');
console.log(result.response);
console.log(`Used ${result.toolCalls.length} tool calls`);

// With abort signal
const controller = new AbortController();
const result = await agent.run('Refactor this file', {
  signal: controller.signal,
});

runParallelSubAgents()

runParallelSubAgents(tasks, options?): Promise<SubAgentResult[]>;

Defined in: agent.ts:2644

Run multiple sub-agents in parallel with proper state isolation.

This method ensures that parallel sub-agents don’t share mutable state, preventing race conditions and state leakage between concurrent runs.

Inspired by LangGraph issue #6446: Parallel subgraphs with shared state keys cause InvalidUpdateError.

Parameters

Parameter	Type	Description
tasks	{ name: string; task: string; }[]	Array of {name, task} objects to run in parallel
options?	RunOptions	Optional run options applied to all sub-agents

Returns

Promise<SubAgentResult[]>

Array of results in the same order as tasks

Example

// Run code review and security scan in parallel
const results = await agent.runParallelSubAgents([
  { name: 'code-reviewer', task: 'Review src/auth.ts' },
  { name: 'security-scanner', task: 'Scan src/auth.ts for vulnerabilities' },
]);

runSubAgent()

runSubAgent(
   name, 
   task, 
options?): Promise<SubAgentResult>;

Defined in: agent.ts:2533

Run a sub-agent with a specific task.

The sub-agent executes independently with its own context and returns the result to the parent agent.

Parameters

Parameter	Type	Description
name	string	Name of the registered sub-agent
task	string	Task description for the sub-agent
options?	RunOptions	Optional run options

Returns

Promise<SubAgentResult>

serialize()

serialize(): AgentState;

Defined in: agent.ts:2245

Serialize the current agent state to an AgentState object. This can be used for manual persistence or transferring state.

Returns

AgentState

Example

const state = agent.serialize();
const json = JSON.stringify(state);
// Store json somewhere...

setHistory()

setHistory(messages, options?): Promise<Agent>;

Defined in: agent.ts:1898

Set the conversation history (for manual compaction/restoration) Also updates the context manager’s token count if configured.

Parameters

Parameter	Type	Description
messages	Message[]	The message history to restore
options?	{ turnCount?: number; }	Optional restore options
options.turnCount?	number	The turn count to restore (important for compaction)

Returns

Promise<Agent>

setModel()

setModel(modelId): void;

Defined in: agent.ts:1581

Change the model for subsequent LLM calls (same provider only).

Takes effect on the next run() or stream() call — never interrupts a running turn. Conversation history is preserved (it’s provider-agnostic). Emits a model_changed event.

Use this to switch between models within the same provider, e.g., Claude Sonnet → Claude Opus for a harder task, then back.

Parameters

Parameter	Type	Description
modelId	string	The new model ID (e.g., ‘claude-opus-4-20250514’)

Returns

void

Throws

If modelId is empty or not a string

Example

console.log(agent.getModel()); // 'claude-sonnet-4-20250514'
agent.setModel('claude-opus-4-20250514');
// Next run() uses Opus
const result = await agent.run('Solve this complex problem');
agent.setModel('claude-sonnet-4-20250514'); // Switch back

Since

0.5.8

setPermissionLevel()

setPermissionLevel(
   toolName, 
   level, 
   description?): this;

Defined in: agent.ts:1760

Set the permission level for a tool

Parameters

Parameter	Type	Description
toolName	string	Tool name or pattern
level	PermissionLevel	Permission level
description?	string	Optional description

Returns

this

stream()

stream(userMessage, options?): AsyncIterable<AgentEvent>;

Defined in: agent.ts:3970

Stream the agent’s response as events.

Yields AgentEvent objects in real time as the agent thinks, calls tools, and generates text. Use this for building interactive UIs that show progress as it happens.

Parameters

Parameter	Type	Description
userMessage	string	The user’s message
options?	RunOptions	Optional run configuration

Returns

AsyncIterable<AgentEvent>

An async iterable of agent events

Example

for await (const event of agent.stream('Explain this code')) {
  if (event.type === 'llm_chunk') {
    process.stdout.write(event.chunk.text ?? '');
  } else if (event.type === 'tool_start') {
    console.log(`\nCalling tool: ${event.name}`);
  } else if (event.type === 'done') {
    console.log('\n\nDone!');
  }
}