Human Engagement ​

When we talk about agents, we usually mean it to be LLM-driven. However, in some real scenarios, we do want humans to step in, whether it's for quality assurance in a project, guidance offering in critical decision-making, or role playing in a game. In this tutorial, we talk about how to take humans into the loop.

After this tutorial, you will be able to:

• Introduce interaction between LLM-based agents and human

Switch between LLM and human ​

We will reuse the exact example in MultiAgent 101.

Originally, an LLM assumes the SimpleReviewer. Suppose we want more control over the review process, we may take up the Role ourselves. This is enabled by just one switch: set is_human=True upon initialization. The code becomes:

team.hire(
    [
        SimpleCoder(),
        SimpleTester(),
        # SimpleReviewer(), # the original line
        SimpleReviewer(is_human=True), # change to this line
    ]
)

team.hire(
    [
        SimpleCoder(),
        SimpleTester(),
        # SimpleReviewer(), # the original line
        SimpleReviewer(is_human=True), # change to this line
    ]
)

We are a human SimpleReviewer interacting with the two LLM-based agents SimpleCoder and SimpleTester now. We can comment on the unit tests from SimpleTester, requesting more coverage or edge cases. The feedback is then sent back to the tester for writing a new version. The switch is complete agnostic to the original SOP and Role definition, meaning applicability to general scenarios.

Each time when it's our turn to respond, the running process will pause to wait for our input. Just type in what we want, we are sending our messages to the agents!

Constraint:
For develops who customize Role's _act function, the Actions called in _act must be among the self._actions initialized with set_actions for the human engagement to take effect.

LIMITATION:
Currently the interaction is through terminal input, which is inconvenient for multi-line or structured writeup. Meanwhile, users must adhere to the prompt in content or format like what we require of an LLM, in order for the logic after human input works as usual. We will provide solutions to these issues in following update.

Complete script of this tutorial ​

https://github.com/geekan/MetaGPT/blob/main/examples/build_customized_multi_agents.py

Run it with

python3 examples/build_customized_multi_agents.py --add_human True

python3 examples/build_customized_multi_agents.py --add_human True

A sample interaction