Configurable Model

Switch models per-call or permanently at runtime without recreating the agent

One agent, many models — override the LLM for a single call with config=, or permanently swap it with switch_model(), all while keeping your conversation history intact.

from praisonaiagents import Agent

agent = Agent(name="flex", instructions="Answer with the best model for each task.", model="gpt-4o-mini")
agent.start("Draft a short poem, then switch to a stronger model for critique.")

The user keeps one agent; per-call config= or switch_model() changes the LLM without losing history.

Override vs Switch

Quick Start

1. Override model for a single call

   from praisonaiagents import Agent
   
   agent = Agent(
       name="FlexBot",
       instructions="Answer questions helpfully.",
       model="gpt-4o-mini"
   )
   
   cheap = agent.chat("Summarize this paragraph.")
   
   expensive = agent.chat(
       "Write a detailed technical analysis.",
       config={"model": "gpt-4o"}
   )

   The agent uses gpt-4o-mini by default. Passing config={"model": "gpt-4o"} upgrades just that one call. The next call reverts to gpt-4o-mini.

2. Permanently switch the agent's model

   from praisonaiagents import Agent
   
   agent = Agent(
       name="AdaptiveBot",
       instructions="Help with various tasks.",
       model="gpt-4o-mini"
   )
   
   agent.chat("Quick question: what's 2 + 2?")
   
   agent.switch_model("gpt-4o")
   
   agent.chat("Now write a 500-word essay on climate change.")

   switch_model() updates the agent in place. All future calls use the new model, and the conversation history is fully preserved.

How It Works

Per-call config overrides are isolated to that single invocation and never mutate the agent’s defaults. This makes them safe for concurrent use — multiple threads can call the same agent with different models simultaneously.

switch_model() updates agent.llm and recreates the internal LLM instance. Conversation history stored in agent._chat_history is untouched.

Configuration Options

chat() config keys:

response = agent.chat(
    "Write a creative story.",
    config={
        "model": "gpt-4o",
        "temperature": 0.9,
    }
)

switch_model() signature:

agent.switch_model("gpt-4o")

Accepts any model string supported by LiteLLM (same format as the llm= constructor argument).

Common Patterns

Route by task complexity:

from praisonaiagents import Agent

agent = Agent(name="Router", instructions="Answer questions.", model="gpt-4o-mini")

def ask(prompt: str, complex: bool = False) -> str:
    if complex:
        return agent.chat(prompt, config={"model": "gpt-4o"})
    return agent.chat(prompt)

simple = ask("What is 10 * 5?")
detailed = ask("Explain quantum entanglement.", complex=True)

Override temperature for creative tasks:

agent = Agent(name="Writer", instructions="Write content.", model="gpt-4o-mini")

factual = agent.chat("What year was Python created?")

creative = agent.chat(
    "Write a haiku about debugging.",
    config={"temperature": 1.2}
)

Escalate on failure:

def safe_chat(agent, prompt):
    try:
        return agent.chat(prompt)
    except Exception:
        return agent.chat(prompt, config={"model": "gpt-4o"})

Best Practices

Related