curiositech/wei-2022-chain-of-thought

Chain-of-Thought Reasoning for Agent Systems

Strategic framework for eliciting and orchestrating reasoning in LLM-based agents through structured decomposition

Decision Points

Primary Routing Decision: Decompose vs. Direct

Task Complexity × Agent Capability → Action

IF agent_parameters < 100B AND task_requires_multi_step_reasoning:
  → DIRECT prompting (decomposition hurts below emergence threshold)
  
ELIF baseline_accuracy > 80% AND reasoning_steps < 3:
  → DIRECT prompting (minimal gains, added latency cost)
  
ELIF task_complexity = HIGH AND multi_step_reasoning = TRUE:
  → CHAIN_OF_THOUGHT prompting (gains scale with complexity)
  
ELIF task_distribution = MIXED_COMPLEXITY:
  → SELF_SELECT mode (provide examples, let agent choose when to decompose)

Failure Classification Decision Tree

Observe failure type:

IF error_type IN [calculator_mistake, symbol_mapping_error, missing_single_step]:
  → SHALLOW failure
  → Add external tools (calculator, validator)
  → Implement consistency checks
  → Re-run with augmentation
  
ELIF error_type IN [semantic_misunderstanding, incoherent_logic, fundamental_reasoning_flaw]:
  → DEEP failure
  → Check agent capability threshold
  → Re-route to more capable agent OR reject task
  
ELIF correct_answer = TRUE:
  → Verify reasoning chain (98% of correct answers have sound reasoning)
  → High confidence in output quality

Reasoning Structure Selection Matrix

Task Type × Semantic Requirements → Structure

IF task = SYMBOLIC_MANIPULATION:
  → HYBRID: Natural language setup → formal operations → natural language verification
  
ELIF semantic_understanding_across_steps = TRUE:
  → FULL natural language chain-of-thought
  → Don't optimize for brevity (semantic grounding requires explicit articulation)
  
ELIF multi_agent_coordination = TRUE:
  → Natural language for coordination medium
  → Formal protocols within individual agents only

Failure Modes

1. Premature Decomposition Syndrome

Detection Rule: If small model (<100B params) + chain-of-thought prompting → worse performance than baseline Symptom: Fluent but illogical reasoning chains, performance degradation Fix: Verify emergence threshold empirically before deploying decomposition; use direct prompting below threshold

2. Raw Compute Fallacy

Detection Rule: If adding compute time/tokens without structured intermediate steps → no performance gain Symptom: Longer outputs with dots/padding but same accuracy Fix: Structure compute through meaningful semantic intermediate states, not just duration

3. Uniform Error Handling

Detection Rule: If same recovery strategy applied to calculator errors and semantic failures Symptom: Tool augmentation fails on deep reasoning errors; model re-routing wastes cycles on shallow errors Fix: Classify failures first (shallow vs. deep), then route to appropriate intervention

4. Optimization Premature Optimization

Detection Rule: If natural language reasoning shortened to save tokens → coherence loss Symptom: Broken semantic grounding, reasoning chain loses logical connection Fix: Preserve explicit articulation; the "inefficiency" maintains semantic coherence

5. Brittleness Blindness

Detection Rule: If demo performance (95%) ≫ production performance (70%) across task variations Symptom: High variance across annotators/exemplar sets on low-complexity tasks Fix: Test robustness envelope before production; expect brittleness on tasks with high baseline accuracy

Worked Examples

Example 1: Math Word Problem Routing Decision

Scenario: Multi-step arithmetic problem arrives at LaMDA 68B agent

Decision Process:

1. Check capability: 68B > 100B threshold? No, but close to emergence
2. Check task complexity: 4 reasoning steps required = HIGH
3. Check baseline accuracy: ~6% without chain-of-thought
4. Decision: Use chain-of-thought (expected jump from 6% to ~14%)

Execution:

• Prompt: "Let's think step by step. [problem]"
• Agent produces: "First I need to find... Then I calculate... Next I..."
• Expert catches: Verify each calculation step has semantic grounding
• Novice misses: Would skip intermediate verification, trust fluent output

Result: 14% accuracy vs. 6% baseline (emergence threshold effect)

Example 2: Failure Classification and Recovery

Scenario: Agent produces wrong answer: "The total cost is $47" (correct: $52)

Analysis Process:

1. Examine reasoning chain: "First, 3 items × $12 = $36. Then add $15 tax = $51. Wait, I mean $47."
2. Classify failure: Calculator error in final step (36 + 15 = 51, but output 47)
3. Failure type: SHALLOW (arithmetic mistake, reasoning structure sound)
4. Decision: Add calculator tool, re-run

Recovery:

• Augment with external calculator
• Re-run same reasoning chain with tool verification
• Expert insight: Don't re-route to bigger model for arithmetic errors
• Novice error: Would assume "model isn't smart enough" and upgrade unnecessarily

Example 3: Natural Language vs. Formal Protocol Choice

Scenario: Multi-agent system coordinating complex financial analysis

Decision Point: Use formal API calls or natural language coordination?

Analysis:

• Task requires semantic understanding across 6 reasoning steps
• Each step builds on previous semantic context
• Formal protocol would lose grounding between symbolic operations and real-world meaning

Decision: Natural language coordination despite token overhead

Implementation:

Agent A: "Given the Q3 earnings show 15% revenue growth but 8% margin compression, 
         I need to analyze if this indicates sustainable growth or pricing pressure..."
Agent B: "Building on your margin analysis, the compression aligns with our competitive 
         positioning data showing 3 new market entrants..."

Expert reasoning: The "inefficiency" of natural language maintains semantic coherence Novice mistake: Would optimize for concise formal protocols and lose reasoning grounding

Example 4: Production Robustness Testing

Scenario: Prompting technique achieves 95% accuracy on test exemplars

Pre-deployment Process:

1. Variance test: Run across 5 different annotator sets
2. Results: Performance range 71% - 96% (high variance = brittleness indicator)
3. Task analysis: Coin flip reasoning (low complexity, high baseline)
4. Decision: Brittle application, implement stabilization or accept variance

Production Strategy:

• Deploy with variance monitoring
• Flag high-variance outputs for review
• Expert insight: Brittleness correlates with low reasoning complexity
• Novice assumption: Would expect consistent performance based on demo results

Quality Gates

Capability Assessment

• [ ] Agent parameter count verified (>100B for chain-of-thought effectiveness)
• [ ] Emergence threshold confirmed empirically on sample tasks
• [ ] Baseline accuracy measured without decomposition
• [ ] Task complexity categorized (steps required, semantic depth)

Failure Mode Preparation

• [ ] Error classification system implemented (shallow vs. deep detection rules)
• [ ] Tool augmentation ready for shallow failures (calculator, validator)
• [ ] Re-routing paths defined for deep failures
• [ ] Recovery strategy mapped to failure type

Reasoning Structure Validation

• [ ] Natural language vs. formal protocol choice justified
• [ ] Semantic grounding maintained across reasoning steps
• [ ] Intermediate states contain meaningful semantic content
• [ ] Brevity optimization balanced against coherence requirements

Production Readiness

• [ ] Robustness tested across multiple annotator/exemplar sets
• [ ] Variance envelope mapped for task distribution
• [ ] Brittleness indicators identified and monitored
• [ ] Performance degradation thresholds defined

Quality Assurance

• [ ] Sample outputs manually verified for reasoning chain coherence
• [ ] Correct answers verified to have sound reasoning (98% confidence rule)
• [ ] Edge cases tested at capability boundaries
• [ ] Fallback strategies defined for out-of-distribution tasks

NOT-FOR Boundaries

This skill is NOT for:

• Simple lookup tasks: Use direct retrieval instead of reasoning decomposition
• Tasks with >90% baseline accuracy: Minimal gains don't justify latency overhead
• Sub-100B parameter models on complex reasoning: Use [model-scaling-strategy] to upgrade capability first
• Pure symbolic computation: Use [formal-reasoning-systems] for equation solving without semantic context
• Real-time applications: Decomposition latency incompatible with <100ms response requirements

Delegate to other skills:

• For model selection decisions → Use [capability-threshold-analysis]
• For tool integration strategy → Use [agent-tool-coordination]
• For multi-agent protocol design → Use [agent-communication-patterns]
• For performance optimization → Use [latency-accuracy-tradeoffs]
• For error monitoring systems → Use [agent-observability-patterns]