OODA Loop

Definition

The OODA Loop—Observe, Orient, Decide, Act—is a strategic decision-making cycle originally developed by U.S. Air Force Colonel John Boyd for aerial combat. Boyd’s insight: the combatant who cycles through these phases faster than their opponent gains decisive advantage. The key isn’t raw speed—it’s tempo relative to environmental change.

Boyd’s less-quoted but crucial insight: Orient is everything. The Orient phase is where mental models, context, and prior experience shape how observations become decisions. A faster but poorly-oriented loop loses to a slower but well-oriented one.

In agentic software development, OODA provides the cognitive model for how autonomous agents should behave: continuously cycling through observation, interpretation, planning, and execution.

The Four Phases

1. Observe — Gather information about the current state of the environment
2. Orient — Interpret observations through mental models, context, and constraints
3. Decide — Formulate a specific plan for action based on orientation
4. Act — Execute the plan, producing changes that feed new observations

The loop is continuous. Each Act produces new state, triggering new Observe, and the cycle repeats.

Key Characteristics

Tempo, Not Raw Speed

The strategic value of OODA isn’t speed—it’s cycling faster than the environment changes. In software development, the “environment” is the codebase, requirements, and constraints. An agent that can cycle through OODA before context rot sets in converges on correct solutions.

Orient as the Critical Phase

For AI agents, Orient is the context window. The quality of orientation depends on:

• Spec Clarity — Garbage spec → garbage orientation
• Constitution Directives — Values that shape interpretation
• Context Gates — Filtering noise so orientation isn’t polluted
• Prior State — Git history, progress files, previous learnings

This is why Context Engineering isn’t optional overhead. It’s engineering the Orient phase, which determines whether fast cycling produces progress or noise.

OODA vs. Single-Shot Interactions

Standard LLM interactions are Observe-Act: user provides input, model produces output. No explicit Orient or Decide phase. The model’s “orientation” is implicit in training and whatever context happens to be present.

Agentic workflows make OODA explicit:

Phase	Single-Shot LLM	Agentic Workflow
Observe	User prompt	Instrumented: read files, run tests, check logs
Orient	Implicit (training + context)	Engineered: Specs, Constitution, Context Gates
Decide	Implicit	Explicit: agent states plan before acting
Act	Generate response	Verified: external tools confirm success/failure

This explicit structure enables debugging. When an agent fails, you can diagnose which phase broke down:

• Bad Observe? Agent missed relevant information
• Bad Orient? Context was polluted or incomplete
• Bad Decide? Plan was incoherent given good orientation
• Bad Act? Execution failed despite good plan

ASDLC Usage

In ASDLC, OODA explains why cyclic workflows outperform linear pipelines:

OODA Phase	Agent Behavior	ASDLC Component
Observe	Read codebase state, error logs, test results	File state, test output
Orient	Interpret against context and constraints	Context Gates, AGENTS.md
Decide	Formulate implementation plan	PBI decomposition
Act	Write code, run tests, commit	Micro-commits

The Learning Loop is OODA with an explicit “Crystallize” step that improves future Orient phases. Where OODA cycles continuously, Learning Loop captures discoveries into machine-readable context for subsequent agent sessions.

Applied in:

• Context Engineering — The discipline of engineering the Orient phase
• Context Gates — Checkpoints between OODA phases
• Levels of Autonomy — Higher autonomy requires more sophisticated Orient capabilities

Anti-Patterns

Anti-Pattern	Description	Failure Mode
Observe-Act	Skipping Orient/Decide. Classic vibe coding.	Works for simple tasks; fails at scale; no learning
Orient Paralysis	Over-engineering context, never acting	Analysis paralysis; no forward progress
Stale Orient	Not updating mental model when observations change	Context rot; agent operates on outdated assumptions
Observe Blindness	Not instrumenting observation of relevant state	Agent misses critical information (failed tests, error logs)
Act Without Verify	Not confirming action results before next cycle	Cascading errors; false confidence