Give the decision
a structure.
Conflicting signals. Shifting conditions. The wrong incentives. Four mental models I use to make the next move clearer.
Find your starting pointOne original framework. Three established models.
Practical adaptations, with their limits made explicit.
Where is the decision getting stuck?
InDecision
Knox’s market-analysis framework
Make the case against your own conviction.
InDecision structures market analysis around competing bullish and bearish evidence, with risk context shaping whether to act. The useful habit travels beyond markets: make the opposing case explicit before committing.
- 01GatherRecord the evidence
- 02CompareBuild both cases
- 03GateCheck risk and limits
- 04DecideAct, wait, or investigate
What evidence would make the opposite conclusion more credible?
A written decision, its strongest counterargument, and the condition that would change it.
Illustrative adaptation · launch decision
A product team is debating a launch. The positive case is strong activation in a small pilot; the negative case is weak repeat use. Instead of averaging those into a confident yes, the team runs a retention check and defines what result would justify a wider release.
Weights and thresholds can encode bias. This overview makes no claim about trading accuracy or returns; a performance claim needs a defined sample, time window, and evaluation method.
OODA Loop
John Boyd’s decision model
Update your picture before your next move.
Observe, Orient, Decide, Act. Orientation is where experience, assumptions, and new information shape your interpretation. Use feedback to revise that picture. Moving quickly with a stale model can repeat the same mistake faster.
- 01ObserveWhat changed?
- 02OrientWhat does it mean?
- 03DecideChoose a bounded move
- 04ActCheck the result
Which assumption changed, and what is the smallest useful action now?
A bounded action, an observable result, and a trigger to reassess.
Illustrative application · agent evaluation
An agent’s test pass rate drops after a model change. Observe the failing cases, separate prompt failures from tool failures, choose a small fix, and rerun the same evaluation. Feed the result into the next decision before expanding the rollout.
Tempo matters alongside judgment, resources, and context. Faster action alone does not guarantee a better outcome.
Principal–Agent Problem
Economics and contract theory
Define success before you delegate authority.
When someone acts on your behalf, their incentives and information may differ from yours. Clear agreements, appropriate incentives, visibility, and accountability help manage that gap. No single metric captures every part of good work.
- 01IntentThe result you need
- 02IncentiveWhat gets rewarded
- 03BehaviorWhat actually happens
- 04ReviewCompare with intent
Could someone hit this target while making the actual outcome worse?
An outcome measure, a quality constraint, and a clear review boundary.
Illustrative analogy · delegated coding
An agent rewarded only for closing tickets may produce shallow fixes. Add acceptance criteria, regression checks, and human approval for sensitive changes. Review whether the underlying problem is resolved, rather than relying on the ticket count.
This economic model concerns people and institutions. Applying it to AI agents is an analogy about objectives and evaluation, not a claim that software has human motives.
Center of Gravity
Carl von Clausewitz · On War
Find the source of strength holding the system together.
Clausewitz used the center of gravity to focus military effort on a source of an opponent’s strength and cohesion. In systems work, I use it as a prompt to examine dependencies: which capability is essential to the result we want?
- 01OutcomeName the objective
- 02DependenciesMap what supports it
- 03TestChallenge the key link
- 04FocusDirect the effort
If this capability disappeared, could the system still deliver the outcome?
A dependency hypothesis and a test that could disprove it.
Illustrative adaptation · knowledge quality
A team keeps improving prompts while its agents retrieve outdated instructions. Map the delivery dependencies and test retrieval quality. If correct context changes the result, fixing the knowledge source may matter more than another prompt revision.
Complex systems can have several sources of strength and can adapt. Do not assume that one intervention will collapse or repair the whole system.
Try it on a decision you’re already making.
Choose one framework. Write the decision, the evidence, and what would change your mind. After acting, compare the result with what you expected.