Reasoning Strategies

TraitClaw supports three built-in reasoning strategies via the AgentStrategy trait, plus the ability to build custom strategies.

Default Strategy

By default, agents use DefaultStrategy — a simple loop of generate → tool call → generate until the LLM returns a text response.

ReAct (Reasoning + Acting)

ReAct interleaves thinking and acting in a structured loop:

Think → Act → Observe → Think → Act → Observe → ... → Answer

use traitclaw_strategies::ReActStrategy;

let strategy = ReActStrategy::builder()
    .max_steps(10)
    .build();

let agent = Agent::builder()
    .provider(provider)
    .strategy(strategy)
    .tool(WebSearch)
    .tool(Calculator)
    .build()?;

Chain-of-Thought (CoT)

CoT forces the LLM to show its reasoning step by step:

use traitclaw_strategies::CotStrategy;

let strategy = CotStrategy::builder()
    .max_steps(5)
    .build();

Monte Carlo Tree Search (MCTS)

MCTS explores multiple reasoning paths in parallel and selects the best:

use traitclaw_strategies::MctsStrategy;

let strategy = MctsStrategy::builder()
    .num_simulations(8)
    .exploration_weight(1.4)
    .build();

Custom Strategy

Implement AgentStrategy for your own reasoning loop:

use traitclaw_core::{AgentStrategy, AgentRuntime, AgentOutput};

#[async_trait]
impl AgentStrategy for MyStrategy {
    async fn execute(
        &self,
        runtime: &AgentRuntime,
        input: &str,
    ) -> Result<AgentOutput> {
        // Your custom reasoning loop
        let output = runtime.generate(input).await?;
        Ok(output)
    }
}