Skip to main content
Agent Type: Background Confidence AssessmentTools: Read, Glob, Grep, BashSpecial: Runs in background with isolated worktree

Overview

The Scout agent performs confidence-gated exploration to assess whether you have enough context to implement a task successfully. It scores readiness 0-100 across five dimensions and delivers a GO/HOLD verdict. Scout runs in the background, allowing you to continue working while it explores.

When to Use

Use Scout before starting implementation of:

Unfamiliar Code

Working in parts of the codebase you haven’t touched before

Complex Tasks

Tasks with unclear scope or many unknowns

High Risk Changes

Changes that could break existing functionality

New Features

Building features without clear patterns to follow

Configuration

---
name: scout
description: Confidence-gated exploration that assesses readiness before implementation. Scores 0-100 across five dimensions and gives GO/HOLD verdict.
tools: ["Read", "Glob", "Grep", "Bash"]
background: true
isolation: worktree
---

Special Capabilities

background
boolean
default:true
Scout runs in the background, allowing you to continue working while it explores the codebase.
isolation
string
default:"worktree"
Scout runs in an isolated worktree to avoid interfering with your main work session.
Scout is read-only. It never edits files or makes changes.

Workflow

Scout follows a confidence-building workflow:
1

Receive Task

Get the task description from you
2

Explore Codebase

Search for relevant files, patterns, and dependencies
3

Score Confidence

Rate readiness 0-100 across five dimensions
4

Verdict

If score >= 70: GO with findingsIf score < 70: Identify gaps, gather more context, re-score

Confidence Scoring

Scout evaluates five dimensions, each worth 0-20 points:

1. Scope Clarity (0-20)

  • Know exactly which files need changes
  • Understand all modifications required
  • Can enumerate every change needed
  • Know most files that need changes
  • Some uncertainty about edge cases
  • Can estimate change scope
  • Unsure which files to modify
  • Don’t understand the full scope
  • Many unknowns remain

2. Pattern Familiarity (0-20)

  • Codebase has similar features to follow
  • Patterns are consistent and documented
  • Can copy existing approach
  • Found some related code
  • Patterns exist but inconsistent
  • Need to adapt existing patterns
  • No similar features exist
  • First time implementing this type of feature
  • Need to create new patterns

3. Dependency Awareness (0-20)

  • Know all dependencies and dependents
  • Understand impact of changes
  • Can predict breaking changes
  • Know major dependencies
  • Some unknowns about impact
  • Can identify most breaking changes
  • Don’t know what depends on this code
  • Can’t predict impact
  • High risk of breaking things

4. Edge Case Coverage (0-20)

  • Can enumerate all edge cases
  • Know how to handle errors
  • Understand validation requirements
  • Know major edge cases
  • Some uncertainty about errors
  • Basic validation understood
  • Don’t know edge cases
  • Unsure how to handle errors
  • Validation requirements unclear

5. Test Strategy (0-20)

  • Know exactly how to test changes
  • Test patterns exist to follow
  • Can verify all functionality
  • Have general testing approach
  • Some test patterns available
  • Can verify main functionality
  • Don’t know how to test
  • No test patterns to follow
  • Unclear how to verify changes

Output Format

SCOUT REPORT
Task: Add user profile page with avatar upload
Confidence: 85/100

Dimensions:
  Scope clarity:        18/20  ✓
  Pattern familiarity:  17/20  ✓
  Dependency awareness: 16/20  ✓
  Edge case coverage:   17/20  ✓
  Test strategy:        17/20  ✓

VERDICT: GO

Findings:
- User pages follow pattern in src/pages/user/*.tsx
- Avatar upload similar to src/components/ImageUpload.tsx
- Profile API endpoints exist: GET/PUT /api/user/:id/profile
- Tests follow pattern in tests/pages/*.test.tsx
- Need to handle: missing avatar, invalid image types, file size limits

Ready to implement. Follow ImageUpload pattern for avatar handling.

Example Scenarios

Task: Add dark mode toggle to settings

SCOUT REPORT
Confidence: 92/100

Dimensions:
  Scope clarity:        19/20  ✓
  Pattern familiarity:  20/20  ✓
  Dependency awareness: 18/20  ✓
  Edge case coverage:   18/20  ✓
  Test strategy:        17/20  ✓

VERDICT: GO

Findings:
- Theme system exists in src/theme/
- Settings page at src/pages/Settings.tsx
- Dark mode CSS variables defined
- Toggle component in src/components/Toggle.tsx
- Tests in tests/theme/*.test.tsx

Files to change:
1. src/pages/Settings.tsx - Add toggle
2. src/hooks/useTheme.ts - Add dark mode state
3. tests/pages/Settings.test.tsx - Add tests

Ready to implement.

Rules and Constraints

Scout adheres to strict rules:
Scout performs read-only exploration only. It never modifies the codebase.
A false GO wastes more time than a HOLD. Scout is brutally honest about confidence.
If score < 70, Scout gathers more context and re-scores. Maximum 2 rounds before escalating.
Runs in isolated worktree to avoid interfering with your main work session.

Best Practices

1

Use Early

Run Scout before starting implementation, not when you’re already stuck
2

Trust the Verdict

If Scout says HOLD, gather more context before proceeding
3

Continue Working

Scout runs in background—keep working on other tasks while it explores
4

Review Findings

Even with GO verdict, review Scout’s findings to inform your approach

Comparison with Planner

FeatureScoutPlanner
OutputGO/HOLD verdictDetailed plan
Scoring0-100 confidenceNot applicable
ExecutionBackgroundForeground
IsolationWorktreeNone
Re-attemptsAuto (max 2)Manual
Use caseQuick confidence checkDetailed planning
Use Scout for quick checks and Planner for detailed architectural planning.

Integration Example

Troubleshooting

Scout runs in background and should not block your work. If it seems slow, check:
  • Very large codebase? May need more time
  • Complex dependencies? Requires thorough exploration
  • Wait for verdict or continue with other tasks
If Scout consistently returns HOLD:
  • Task might be too complex for immediate implementation
  • Codebase might lack patterns for this feature
  • Consider breaking task into smaller pieces
  • Use Planner for architectural guidance
If Scout says GO but you’re uncertain:
  • Review the findings section carefully
  • Score may be borderline (70-75)
  • Run Planner for more detailed guidance
  • Trust your judgment—human intuition matters

Next Steps

Orchestrator

Multi-phase implementation after GO verdict

Planner

Detailed planning for HOLD scenarios

Build docs developers (and LLMs) love