Agent API

The Agent class is the core component of Tyler, responsible for managing conversations and executing tasks. It uses Pydantic for data validation and Weave for monitoring and tracing.

Initialization

from tyler.models.agent import Agent

agent = Agent(
    model_name: str = "gpt-4o",
    temperature: float = 0.7,
    name: str = "Tyler",
    purpose: str = "To be a helpful assistant.",
    notes: str = "",
    tools: List[Union[str, Dict]] = [],
    max_tool_iterations: int = 10,
    thread_store: Optional[ThreadStore] = None
)

Parameters

Parameter	Type	Required	Default	Description
model_name	str	No	"gpt-4o"	The LLM model to use
temperature	float	No	0.7	Response creativity (0.0-1.0)
name	str	No	"Tyler"	Name of the agent
purpose	str	No	"To be a helpful assistant."	The agent's purpose or role
notes	str	No	""	Additional notes for the agent
tools	List[Union[str, Dict]]	No	[]	List of tools (strings for built-in modules or dicts for custom tools)
max_tool_iterations	int	No	10	Maximum number of tool iterations
thread_store	Optional[ThreadStore]	No	ThreadStore()	Thread storage implementation

Methods

go

Process a thread and generate responses.

async def go(
    self,
    thread_or_id: Union[str, Thread],
    new_messages: Optional[List[Message]] = None
) -> Tuple[Thread, List[Message]]

Parameters

Parameter	Type	Required	Description
thread_or_id	Union[str, Thread]	Yes	Thread object or thread ID to process
new_messages	Optional[List[Message]]	No	Messages added during this processing round

Returns

Type	Description
Tuple[Thread, List[Message]]	The processed thread and list of new non-user messages

Example

thread = Thread()
message = Message(role="user", content="Hello!")
thread.add_message(message)

processed_thread, new_messages = await agent.go(thread)

go_stream

Process a thread with streaming updates. This method provides real-time updates as the agent processes the request.

async def go_stream(
    self,
    thread: Thread
) -> AsyncGenerator[StreamUpdate, None]

Parameters

Parameter	Type	Required	Description
thread	Thread	Yes	Thread object to process

Returns

An async generator that yields StreamUpdate objects containing:

• Content chunks as they arrive
• Complete assistant messages with tool calls
• Tool execution results
• Final thread state
• Any errors that occur

StreamUpdate Types

Type	Description	Data Content
CONTENT_CHUNK	Partial content from assistant	str
ASSISTANT_MESSAGE	Complete assistant message	Message
TOOL_MESSAGE	Tool execution result	Message
COMPLETE	Final thread state	Tuple[Thread, List[Message]]
ERROR	Error during processing	str

Example

thread = Thread()
message = Message(role="user", content="Hello!")
thread.add_message(message)

print("Assistant: ", end="", flush=True)
async for update in agent.go_stream(thread):
    if update.type == StreamUpdate.Type.CONTENT_CHUNK:
        # Print content chunks as they arrive
        print(update.data, end="", flush=True)
    elif update.type == StreamUpdate.Type.TOOL_MESSAGE:
        # Print tool results on new lines
        tool_message = update.data
        print(f"\nTool ({tool_message.name}): {tool_message.content}")
    elif update.type == StreamUpdate.Type.ERROR:
        # Print any errors that occur
        print(f"\nError: {update.data}")
    elif update.type == StreamUpdate.Type.COMPLETE:
        # Final update contains (thread, new_messages)
        print()  # Add newline after completion

step

Execute a single step of the agent's processing.

@weave.op()
async def step(
    self,
    thread: Thread,
    stream: bool = False
) -> Tuple[Any, Dict]

Parameters

Parameter	Type	Required	Description
thread	Thread	Yes	The thread to process
stream	bool	No	Whether to stream the response

Returns

Type	Description
Tuple[Any, Dict]	The completion response and metrics

Private Methods

_get_completion

Get a completion from the LLM with weave tracing.

@weave.op()
async def _get_completion(
    self,
    **completion_params
) -> Any

_get_thread

Get thread object from ID or return the thread object directly.

@weave.op()
async def _get_thread(
    self,
    thread_or_id: Union[str, Thread]
) -> Thread

_serialize_tool_calls

Serialize tool calls to a list of dictionaries.

@weave.op()
def _serialize_tool_calls(
    self,
    tool_calls: Optional[List[Any]]
) -> Optional[List[Dict]]

_process_tool_call

Process a single tool call and return whether to break the iteration.

@weave.op()
async def _process_tool_call(
    self,
    tool_call,
    thread: Thread,
    new_messages: List[Message]
) -> bool

_handle_max_iterations

Handle the case when max iterations is reached.

@weave.op()
async def _handle_max_iterations(
    self,
    thread: Thread,
    new_messages: List[Message]
) -> Tuple[Thread, List[Message]]

_handle_tool_execution

Execute a single tool call and format the result message.

@weave.op()
async def _handle_tool_execution(
    self,
    tool_call
) -> dict

_normalize_tool_call

Convert a tool_call dict to an object with attributes so it can be used by tool_runner.

@weave.op()
def _normalize_tool_call(
    self,
    tool_call
)

_process_streaming_chunks

Process streaming chunks from the LLM.

@weave.op()
async def _process_streaming_chunks(
    self,
    chunks
) -> Tuple[str, str, List[Dict], Dict]

Private Attributes

Attribute	Type	Description
_prompt	AgentPrompt	Handles system prompt generation
_iteration_count	int	Tracks tool iteration count
_processed_tools	List[Dict]	Stores processed tool definitions

StreamUpdate Class

The StreamUpdate class is used to represent updates from streaming responses.

class StreamUpdate:
    class Type(Enum):
        CONTENT_CHUNK = "content_chunk"      # Partial content from assistant
        ASSISTANT_MESSAGE = "assistant_message"  # Complete assistant message with tool calls
        TOOL_MESSAGE = "tool_message"        # Tool execution result
        COMPLETE = "complete"                # Final thread state and messages
        ERROR = "error"                      # Error during processing
        
    def __init__(self, type: Type, data: Any):
        self.type = type
        self.data = data

Error Handling

try:
    processed_thread, new_messages = await agent.go(thread)
except ValueError as e:
    # Handle validation errors (e.g., thread not found)
    print(f"Validation error: {e}")
except Exception as e:
    # Handle other errors
    print(f"Error: {e}")

Tool Configuration

Built-in Tool Modules

agent = Agent(
    tools=[
        "web",      # Use built-in web tools
        "slack",    # Use built-in Slack tools
        "notion",   # Use built-in Notion tools
        "image",    # Use built-in image tools
        "command_line"  # Use built-in command line tools
    ]
)

Custom Tools

custom_tool = {
    "definition": {
        "type": "function",
        "function": {
            "name": "custom_tool",
            "description": "Tool description",
            "parameters": {
                "type": "object",
                "properties": {
                    "param1": {"type": "string"}
                }
            }
        }
    },
    "implementation": lambda param1: f"Result: {param1}",
    "attributes": {
        "type": "standard"  # or "interrupt"
    }
}

agent = Agent(tools=[custom_tool])

Best Practices

1. Thread Store Selection

   # For development/testing
   agent = Agent()  # Uses in-memory store
   
   # For production
   from tyler.database.thread_store import ThreadStore
   store = ThreadStore("postgresql://...")
   agent = Agent(thread_store=store)

2. Tool Iteration Management

   # For complex workflows
   agent = Agent(max_tool_iterations=20)
   
   # For simple tasks
   agent = Agent(max_tool_iterations=5)

3. Monitoring with Weave

   # Enable Weave tracing
   import weave
   weave.init("tyler")
   
   # Traces will be available for step() and go() methods

See Also

• Thread API
• Message API
• Examples