Utility Skill
Overview
A decision framework for agent orchestration based on Liu et al.,
"Utility-Guided Agent Orchestration for Efficient LLM Tool Use"
(arXiv:2603.19896).
Each candidate action is scored by subtracting weighted costs from
expected gain, producing a single utility value that guides action
selection.
The framework prevents over-calling tools and premature stopping by
making both errors costly.
Utility range is [-2.3, 1.0].
When To Use
- Deciding whether to dispatch another agent or tool call
- Gating expensive tool calls (search, code execution, delegation)
- Selecting the right model tier for a sub-task
- Continuation decisions after receiving partial results
- Verification gating before writing or committing output
When NOT to Use
- Single-step operations with one obvious action
- Trivial tasks where cost of scoring exceeds benefit
- Already-committed actions that cannot be undone
Action Space
A = {respond, retrieve, tool_call, verify, delegate, stop}
| Action |
Description |
| respond |
Emit a final answer from current context |
| retrieve |
Fetch additional information (search, read, lookup) |
| tool_call |
Execute a tool (code runner, API, file write) |
| verify |
Check a prior result for correctness or completeness |
| delegate |
Spawn a sub-agent or hand off to a specialist |
| stop |
Terminate the loop and return current state |
Utility Function
U(a | s_t) = Gain(a | s_t)
- λ₁ · StepCost(a | s_t)
- λ₂ · Uncertainty(a | s_t)
- λ₃ · Redundancy(a | s_t)
| Parameter |
Default |
Rationale |
| λ₁ |
1.0 |
Cost baseline; all other weights relative to this |
| λ₂ |
0.5 |
Weak empirical correlation with outcome (r=0.0131) |
| λ₃ |
0.8 |
Redundancy pruning yields ~10% token savings |
Utility range: [-2.3, 1.0].
Positive values indicate the action is worth taking.
Values below the floor (-0.5 default) indicate the action should
be skipped.
Termination Conditions
Stop the loop when any of the following is true:
- (a) Selected action is
stop
- (b) Step budget exhausted (default: 10 steps)
- (c) All non-
stop actions score below the floor (default: -0.5)
High-gain override: If Gain >= 0.7 for any action, condition
(c) may be overridden.
Document the override and the gain value in your reasoning trace.
Quick Start
Minimal 4-step advisory pattern:
- Construct state: gather task context per
modules/state-builder.md
- Score candidates: evaluate each action in
A per
modules/action-selector.md
- Prefer highest utility: select the action with the
maximum
U(a | s_t), subject to termination conditions
- Log score and decision: record the winning action,
its utility value, and step count before executing
Detailed Resources
- State Builder:
modules/state-builder.md, how to
populate s_t from task context
- Gain:
modules/gain.md, estimating expected information
or progress gain
- Step Cost:
modules/step-cost.md, token, latency, and
monetary cost tables
- Uncertainty:
modules/uncertainty.md, confidence
estimation and calibration
- Redundancy:
modules/redundancy.md, detecting duplicate
or low-delta actions
- Action Selector:
modules/action-selector.md, scoring
loop and tie-breaking rules
- Integration:
modules/integration.md, wiring utility
scoring into existing orchestration loops
Exit Criteria
1---2name: utility-23description: Scores agent actions by expected gain, cost, uncertainty, and redundancy. Use when deciding whether to dispatch an agent or invoke a tool.4---5# Utility Skill67## Overview89A decision framework for agent orchestration based on Liu et al.,10"Utility-Guided Agent Orchestration for Efficient LLM Tool Use"11(arXiv:2603.19896).12Each candidate action is scored by subtracting weighted costs from13expected gain, producing a single utility value that guides action14selection.15The framework prevents over-calling tools and premature stopping by16making both errors costly.17Utility range is [-2.3, 1.0].1819## When To Use2021- Deciding whether to dispatch another agent or tool call22- Gating expensive tool calls (search, code execution, delegation)23- Selecting the right model tier for a sub-task24- Continuation decisions after receiving partial results25- Verification gating before writing or committing output2627## When NOT to Use2829- Single-step operations with one obvious action30- Trivial tasks where cost of scoring exceeds benefit31- Already-committed actions that cannot be undone3233## Action Space3435`A = {respond, retrieve, tool_call, verify, delegate, stop}`3637| Action | Description |38|-----------|------------------------------------------------------|39| respond | Emit a final answer from current context |40| retrieve | Fetch additional information (search, read, lookup) |41| tool_call | Execute a tool (code runner, API, file write) |42| verify | Check a prior result for correctness or completeness |43| delegate | Spawn a sub-agent or hand off to a specialist |44| stop | Terminate the loop and return current state |4546## Utility Function4748```49U(a | s_t) = Gain(a | s_t)50 - λ₁ · StepCost(a | s_t)51 - λ₂ · Uncertainty(a | s_t)52 - λ₃ · Redundancy(a | s_t)53```5455| Parameter | Default | Rationale |56|-----------|---------|---------------------------------------------------|57| λ₁ | 1.0 | Cost baseline; all other weights relative to this |58| λ₂ | 0.5 | Weak empirical correlation with outcome (r=0.0131) |59| λ₃ | 0.8 | Redundancy pruning yields ~10% token savings |6061Utility range: **[-2.3, 1.0]**.62Positive values indicate the action is worth taking.63Values below the floor (-0.5 default) indicate the action should64be skipped.6566## Termination Conditions6768Stop the loop when **any** of the following is true:6970- (a) Selected action is `stop`71- (b) Step budget exhausted (default: 10 steps)72- (c) All non-`stop` actions score below the floor (default: -0.5)7374**High-gain override:** If `Gain >= 0.7` for any action, condition75(c) may be overridden.76Document the override and the gain value in your reasoning trace.7778## Quick Start7980Minimal 4-step advisory pattern:81821. **Construct state**: gather task context per83 `modules/state-builder.md`842. **Score candidates**: evaluate each action in `A` per85 `modules/action-selector.md`863. **Prefer highest utility**: select the action with the87 maximum `U(a | s_t)`, subject to termination conditions884. **Log score and decision**: record the winning action,89 its utility value, and step count before executing9091## Detailed Resources9293- **State Builder**: `modules/state-builder.md`, how to94 populate `s_t` from task context95- **Gain**: `modules/gain.md`, estimating expected information96 or progress gain97- **Step Cost**: `modules/step-cost.md`, token, latency, and98 monetary cost tables99- **Uncertainty**: `modules/uncertainty.md`, confidence100 estimation and calibration101- **Redundancy**: `modules/redundancy.md`, detecting duplicate102 or low-delta actions103- **Action Selector**: `modules/action-selector.md`, scoring104 loop and tie-breaking rules105- **Integration**: `modules/integration.md`, wiring utility106 scoring into existing orchestration loops107108## Exit Criteria109110- [ ] State constructed with task goal and prior steps111- [ ] All six actions scored before selecting one112- [ ] Termination condition checked after each step113- [ ] Score and decision logged for each step taken114- [ ] High-gain overrides documented with gain value
athola