TapeAgents is a framework that leverages a structured, replayable log (Tape) of the agent session to facilitate all stages of the LLM Agent development lifecycle. In TapeAgents, the agent reasons by processing the tape and the LLM output to produce new thoughts, actions, control flow steps and append them to the tape. The environment then reacts to the agent’s actions by likewise appending observation steps to the tape.
Key features:
- Build your agent as a low-level state machine, as a high-level multi-agent team configuration, or as a mono-agent guided by multiple prompts
- Debug your agent with TapeAgent studio or TapeBrowser apps
- Serve your agent with response streaming
- Optimize your agent's configuration using successful tapes; finetune the LLM using revised tapes.
The Tape-centric design of TapeAgents will help you at all stages of your project:
- Build with ultimate flexibility of having access to tape for making prompts and generating next steps
- Change your prompts or team structure and resume the debug session as long as the new agent can continue from the older tape
- Fully control the Agent's tape and the Agent's acting when you use a TapeAgent in an app
- Optimize tapes and agents using the carefully crafted metadata structure that links together tapes, steps, llm calls and agent configurations
We highly recommend starting with the introductory Jupyter notebook. The notebook will introduce you to all the core concepts of framework.
- Clone the repository:
git clone https://github.com/ServiceNow/TapeAgents.git
cd TapeAgents
- Create conda environment and install the package in editable mode:
make setup
To run the example, first activate the tapeagents environment that was created during the setup process:
conda activate tapeagentsThe simplest agent just to show the basic structure of the agent:
from tapeagents.agent import Agent, Node
from tapeagents.core import Prompt
from tapeagents.dialog_tape import AssistantStep, UserStep, DialogTape
from tapeagents.llms import LLMStream, LiteLLM
from tapeagents.prompting import tape_to_messages
llm = LiteLLM(model_name="gpt-4o-mini")
class MainNode(Node):
def make_prompt(self, agent: Agent, tape: DialogTape) -> Prompt:
# Render the whole tape into the prompt, each step is converted to message
return Prompt(messages=tape_to_messages(tape))
def generate_steps(self, agent: Agent, tape: DialogTape, llm_stream: LLMStream):
# Generate single tape step from the LLM output messages stream.
yield AssistantStep(content=llm_stream.get_text())
agent = Agent[DialogTape].create(llm, nodes=[MainNode()])
start_tape = DialogTape(steps=[UserStep(content="Tell me about Montreal in 3 sentences")])
final_tape = agent.run(start_tape).get_final_tape() # agent will start executing the first node
print(f"Final tape: {final_tape.model_dump_json(indent=2)}")The examples/ directory contains examples of how to use the TapeAgents framework for building, debugging, serving and improving agents. Each example is a self-contained Python script (or module) that demonstrates how to use the framework to build an agent for a specific task.
- How to build a single agent that does planning, searches the web and uses code interpreter to answer knowledge-grounded questions, solving the tasks from the GAIA benchmark.
- How to build a team of TapeAgents with AutoGen-style low-code programming paradigm
- How to finetune a TapeAgent with a small LLM to be better at math problem solving on GSM-8k dataset.
Other notable examples that demonstrate the main aspects of the framework:
- workarena - custom agent that solves WorkArena benchmark using BrowserGym environment.
- tape_improver.py - the agent that revisit and improves the tapes produced by another agent.
See our technical report on TapeAgents.
Feel free to reach out to the team:
- Dzmitry Bahdanau, [email protected]
- Oleh Shliazhko, [email protected]
- Jordan Prince Tremblay, [email protected]
We acknowledge the inspiration we took from prior frameworks, in particular LangGraph, AutoGen, AIWaves Agents and DSPy.
