Skip to content

Agents

Conventional wisdom holds that "one agent = one LLM". But we believe that an agent is a program. This program happens to use LLMs to automate much of its decision making, but the program can also ask functions and the user.

As a result, agents in ParaLLeM are more powerful.

An agent is just a python function

The most idiomatic way to declare an agent is with a vanilla Python function that takes a pllm.AgentContext. For example:

import parallem as pllm

def syllable_count_agent(agt: pllm.AgentContext, word: str) -> str:
    ct = agt.ask_llm(f'How many syllables are in "{word}"?')
    return ct.final_answer

with pllm.resume_directory(".pllm/example", llm="gpt-5-nano", strategy="sync", load_dotenv=True) as orch:
    with orch.agent() as agt:
        print(syllable_count_agent(agt, "Hello"))

It is just a python function, nothing special!

An agent is more than an LLM

After decoupling the agent concept from one single LLM, it becomes natural to express concepts including:

  1. One agent calling multiple LLMs (dynamic model switching): 

    def polyglot_agent(agt: pllm.AgentContext):
    resp = agt.ask_llm("What is your model name?", llm="gpt-5-nano")
    resp2 = agt.ask_llm("What is your model name?", llm="gemini-2.5-flash")

  2. Non-linear, branching conversations 

    def animal_agent(agt: pllm.AgentContext):
    resp = agt.ask_llm("Please simply name 8 animals in a comma-separated list.")
    for animal in resp.split(","):
        resp2 = agt.ask_llm(f"Please write a paragraph on the {animal.final_answer.strip()}.")

  3. Asking non-LLMs 

    def best_animal_agent(agt: pllm.AgentContext):
    resp = agt.ask_llm("What is the best animal?")
    resp2 = agt.ask_human(f"Would you agree that {resp.final_answer} is the best animal?")

See also