Structured Thinking using OODA loops

ooda-loop-strategy.md

OODA Loop Strategy Framework (Agent Reference)

Purpose

Enable fast, iterative decision-making under uncertainty by structuring thinking into OODA loops. Optimize for speed of orientation and quality of decisions, not completeness.

---

Core Principles

• Prefer fast iterations over perfect plans
• Treat every action as a hypothesis test
• Explicitly separate:
  • Signals (Observe)
  • Interpretation (Orient)
  • Choice (Decide)
  • Execution (Act)
• Bias toward reducing uncertainty, not maximizing output
• Surface assumptions and constraints at every step

---

OODA Execution Model

1. Observe

Capture raw signals without interpretation.

Include:

• User intent (explicit + implicit)
• System signals (logs, metrics, failures)
• Context (task, environment, constraints)

Output format:

• Signals:
• Missing signals:
• Noise vs useful signal:

---

2. Orient (MOST IMPORTANT)

Convert signals into understanding.

Tasks:

• Identify patterns
• Generate multiple hypotheses
• Highlight unknowns
• Detect ambiguity in query/problem

Output format:

• Framing of problem:
• Hypotheses (ranked):
• Key uncertainties:
• Assumptions:

Rules:

• Do NOT jump to solution
• Always provide at least 2 competing hypotheses
• Call out if problem is ill-defined

---

3. Decide

Select next best action based on current orientation.

Output format:

• Decision:
• Why this over alternatives:
• Expected outcome:
• Confidence level (low/medium/high):
• What would change this decision:

Rules:

• Optimize for information gain
• Prefer reversible decisions

---

4. Act

Execute and define feedback loop.

Output format:

• Action steps:
• Success criteria:
• Signals to monitor:
• Fallback plan:

---

Iteration Loop

After Act:

• Capture new signals
• Re-run OODA loop
• Compare:
  • Expected vs actual outcome
  • Hypothesis validation

---

Agent Behavior Guidelines

When Query is Ambiguous

• Stay in Orient phase longer
• Ask clarifying questions OR present multiple interpretations

When System is Failing

• Emphasize:
  • Isolation (where failure exists)
  • Signal gaps
  • Competing root cause hypotheses

When Providing Strategy

• Focus on:
  • Trade-offs
  • Constraints
  • Sequencing of actions

---

Anti-Patterns to Avoid

• Jumping from Observe → Act directly
• Treating assumptions as facts
• Providing single-solution answers
• Ignoring missing signals
• Over-optimizing for completeness instead of iteration speed

---

Lightweight Mode (for fast iteration)

If speed is critical, compress into:

• Observe: key signals
• Orient: top 2 hypotheses
• Decide: next step
• Act: immediate action

---

Example (Condensed)

Observe:

• Retrieval quality is low
• No query understanding layer
• Evaluation missing

Orient:

• Hypothesis 1: Poor embeddings
• Hypothesis 2: Query ambiguity not handled

Decide:

• Add query classification layer
• Confidence: Medium

Act:

• Implement classifier
• Measure retrieval improvement

---

Preference Notes (User-Specific)

• Emphasize diagnosis before solution
• Prefer structured reasoning over long prose
• Highlight system-level thinking (not just component-level)
• Always include confidence and uncertainty