Bratman Israel Pollack 1988 Plans Resource Bounded Reasoning

Plans and Resource-Bounded Practical Reasoning

Source basis: Bratman, Israel, and Pollack on why plans matter mainly because they constrain future reasoning, not because they fully specify perfect action.

When to Use

• Deliberation time is expensive and the world changes while the agent is still thinking.
• You need to decide when to commit, defer detail, or reconsider a plan.
• A planner is thrashing between options or freezing in analysis paralysis.
• You are designing multi-agent coordination without constant message passing.
• Task decomposition needs to stay executable before all details are known.

NOT for

• Exhaustive planning in toy domains where the world can be treated as frozen.
• Static optimization problems where deliberation cost does not materially matter.
• Flat task queues that do not need commitment structure, revision policy, or means-end coherence.

Decision Points

1. Decide whether a partial commitment now will save more reasoning than it risks in later revision.
2. Choose what stays fixed and what remains open; structural partiality is a design decision, not a temporary defect.
3. Set an override threshold for when incompatible options deserve costly reconsideration.
4. Check whether coordination depends on shared commitment structure or on explicit communication.

Decision Flow

flowchart TD
  A[Planning under resource bounds] --> B{Commit now or deliberate longer?}
  B -->|Commit now| C[Choose partial commitments]
  B -->|Deliberate longer| D[Only if expected value exceeds reasoning cost]
  C --> E[Mark what stays fixed and what remains open]
  D --> E
  E --> F{Override trigger hit?}
  F -->|No| G[Execute with local consistency checks]
  F -->|Yes| H[Reconsider only the threatened commitments]
  G --> I{Coordination needed?}
  H --> I
  I -->|Yes| J[Share commitment structure]
  I -->|No| K[Keep local plan authority]

Working Model

• Plans are computational constraints. They narrow later search, filter incompatible options, and provide assumptions for follow-on reasoning.
• Partiality is a feature. Commit high enough to organize work, but defer details whose stability horizon is short.
• Stability and revisability are in tension. A plan that never filters is useless, and a plan that never bends is brittle.
• Filter override is the meta-reasoning valve. It should trigger reconsideration for meaningful opportunities or failures, not every novelty.
• Means-end coherence turns "what should I do?" into smaller subproblems attached to an already chosen end.

Failure Modes

• Trying to build a complete plan before acting, so the world changes before execution begins.
• Reconsidering so often that commitments never produce computational savings.
• Treating commitments as sacred and missing obvious opportunities or violated assumptions.
• Refining low-level details long before they are needed.
• Ignoring deliberation cost when evaluating whether more planning is worth it.

Anti-Patterns and Shibboleths

• Anti-pattern: treating every new signal as grounds for full replanning.
• Anti-pattern: calling a plan "flexible" when it is really too weak to filter any downstream choices.
• Shibboleth: if the design never prices deliberation cost, it is still doing idealized planning, not bounded practical reasoning.

Worked Examples

• An on-call remediation agent keeps reopening its plan every time a new metric moves. The repair is a stronger filter plus explicit override triggers for only the metrics that threaten the current intention.
• A multi-agent workflow needs independent progress without constant chatter. The repair is to give each agent compatible partial commitments and local consistency checks instead of forcing every step through a central coordinator.

Fork Guidance

• Stay in-process when you are tuning one plan's level of commitment and one override policy.
• Fork only when you need a small number of rival partial plans evaluated independently for expected deliberation cost and failure exposure.

Quality Gates

• Deliberation cost is explicit, not assumed free.
• The plan states which commitments are fixed now and which details remain intentionally open.
• Override triggers are named clearly enough to distinguish useful reconsideration from noise.
• Monitoring covers the assumptions that would invalidate the current plan.
• The design explains how commitments reduce later reasoning load.

Reference Routing

• references/plans-as-computational-constraints.md: load when you need the strongest account of why plans constrain future reasoning.
• references/structural-partiality-and-hierarchical-commitment.md: load when deciding what to commit now versus later.
• references/filter-override-mechanisms.md: load when reconsideration is either too frequent or too rare.
• references/deliberation-as-expensive-operation.md: load when resource budgets, deadlines, or interruption policy matter.
• references/consistency-maintenance-as-coordination-mechanism.md: load when multiple agents need to stay compatible without continuous communication.
• references/temporal-dynamics-of-commitment.md: load when timing of commitment, refinement, or abandonment is the hard part.
• references/failure-modes-of-resource-bounded-reasoning.md: load when diagnosing which bounded-reasoning failure mode you are actually seeing.