TRM (Tiny-Recursive-Model) execution strategy for Jido agents.

This strategy implements recursive reasoning by iteratively improving answers through a reason-supervise-improve cycle. Each iteration:

1. Reasoning: Generate insights about the current answer
2. Supervision: Evaluate the answer and provide feedback
3. Improvement: Apply feedback to generate a better answer

Overview

TRM uses a tiny network applied recursively to iteratively improve answers, achieving remarkable parameter efficiency while outperforming larger models on complex reasoning tasks. Key features:

• Recursive reasoning loop with iterative answer improvement
• Deep supervision with multiple feedback steps
• Adaptive Computational Time (ACT) for early stopping
• Latent state management across recursion steps

Architecture

This strategy uses a pure state machine (Jido.AI.Reasoning.TRM.Machine) for all state transitions. The strategy acts as a thin adapter that:

• Converts instructions to machine messages
• Converts machine directives to SDK-specific directive structs
• Manages the machine state within the agent

Configuration

Configure via strategy options when defining your agent:

use Jido.Agent,
  name: "my_trm_agent",
  strategy: {
    Jido.AI.Reasoning.TRM.Strategy,
    model: "anthropic:claude-sonnet-4-20250514",
    max_supervision_steps: 5,
    act_threshold: 0.9
  }

Options

• :model (optional) - Model alias or direct model spec, defaults to :fast (resolved via Jido.AI.resolve_model/1)
• :max_supervision_steps (optional) - Maximum iterations before termination, defaults to 5
• :act_threshold (optional) - Confidence threshold for early stopping, defaults to 0.9

Signal Routing

This strategy implements signal_routes/1 which AgentServer uses to automatically route these signals to strategy commands:

• "ai.trm.query" → :trm_start
• "ai.llm.response" → :trm_llm_result
• "ai.llm.delta" → :trm_llm_partial

State

State is stored under agent.state.__strategy__ with TRM-specific structure.