Skip to content

strategy_selection_playbook.md

Latest commit

 

History

History
96 lines (75 loc) · 3.48 KB

strategy_selection_playbook.md

File metadata and controls

96 lines (75 loc) · 3.48 KB
Raw

Strategy Selection Playbook

You need to choose a strategy before building agents at scale.

After this guide, you can select CoD, CoT, ReAct, AoT, ToT, GoT, TRM, or Adaptive with explicit tradeoffs.

Strategy Matrix

Strategy Use It For Avoid It For Agent Macro
CoD Token-efficient linear reasoning with concise drafts Deep branching exploration Jido.AI.CoDAgent
CoT Linear reasoning, clear step decomposition Heavy tool orchestration Jido.AI.CoTAgent
ReAct Tool calls + reasoning loop Purely static problems Jido.AI.Agent
AoT One-pass algorithmic exploration with explicit final answer Deep multi-round search orchestration Jido.AI.AoTAgent
ToT Branching search and planning Low-latency simple Q&A Jido.AI.ToTAgent
GoT Multi-perspective synthesis Small deterministic tasks Jido.AI.GoTAgent
TRM Iterative improvement / recursive refinement Fast one-pass answers Jido.AI.TRMAgent
Adaptive Mixed workloads where task type varies Hard real-time deterministic behavior Jido.AI.AdaptiveAgent

Fast Default Recommendation

  • Start with ReAct if tools matter.
  • Start with CoD if reasoning is linear and latency/cost matter.
  • Use CoT when you need more verbose reasoning traces.
  • Start with AoT when you want strict single-query reasoning with explicit answer: extraction.
  • Use Adaptive only when workload shape varies significantly.

Runnable Baseline: Adaptive Agent

defmodule MyApp.SmartAgent do
  use Jido.AI.AdaptiveAgent,
    name: "smart_agent",
    model: :capable,
    default_strategy: :react,
    available_strategies: [:cod, :cot, :react, :tot, :got, :trm]
end

{:ok, pid} = Jido.AgentServer.start(agent: MyApp.SmartAgent)
{:ok, result} = MyApp.SmartAgent.ask_sync(pid, "Compare three migration plans and pick one")

Failure Mode: Wrong Strategy For Task Shape

Symptom:

  • High latency with little quality gain
  • Excess iterations without better answers

Fix:

  • Move from ToT/GoT to CoT for linear problems
  • Move from CoT to CoD for lower latency/cost
  • Move from CoT to ReAct when tools are essential
  • Constrain Adaptive with available_strategies

Defaults You Should Know

  • Adaptive default strategy: :react
  • Adaptive default available strategies: [:cod, :cot, :react, :tot, :got, :trm]
  • Adaptive can include AoT via opt-in list update: available_strategies: [:cod, :cot, :react, :aot, :tot, :got, :trm]
  • AoT defaults: profile: :standard, search_style: :dfs, temperature: 0.0, max_tokens: 2048, require_explicit_answer: true
  • TRM default max_supervision_steps: 5
  • ToT defaults: top_k: 3, min_depth: 2, max_nodes: 100, max_tool_round_trips: 3

ToT Flexible Config (SDK)

defmodule MyApp.PlanningAgent do
  use Jido.AI.ToTAgent,
    name: "planning_agent",
    branching_factor: 3,
    max_depth: 4,
    top_k: 3,
    min_depth: 2,
    max_nodes: 120,
    max_duration_ms: 20_000,
    convergence_window: 2,
    min_score_improvement: 0.02,
    tools: [MyApp.Actions.WeatherLookup],
    max_tool_round_trips: 3
end

When To Use / Not Use

Use this playbook when:

  • You are choosing an architecture for new agents
  • You are triaging quality/latency tradeoffs

Do not use this playbook when:

  • You only need one known strategy already validated in production

Next