Skill Resiliency

Skill Resiliency

Overview

This skill applies Michael Levin's biological resiliency principles to Agent Skills design. It provides a framework for building robust, self-correcting skills that maintain goal-directed behavior despite perturbations, with resiliency mechanisms scaled proportionally to the skill's determinism requirements.

Core Principle: The degree of resiliency in a skill should be positively correlated with its determinism. More deterministic skills (those requiring precise, repeatable outcomes) need stronger resiliency mechanisms to maintain their target behavior.

Inspired by: Michael Levin's research on anatomical homeostasis, error correction in biological systems, and collective intelligence in cellular networks.

When to Use This Skill

Apply resiliency patterns when:

• Creating skills that require deterministic outcomes
• Building skills with complex multi-step workflows
• Designing skills that interact with external systems or tools
• Improving existing skills that fail under edge cases
• Developing skills for critical operations (deployment, data processing, security)

Resiliency Level Guide:

• High Determinism (scripts, automation, data validation) → High Resiliency (extensive validation, recovery, monitoring)
• Medium Determinism (structured workflows, domain guidance) → Medium Resiliency (checkpoint validation, error detection)
• Low Determinism (creative tasks, exploration, ideation) → Low Resiliency (minimal constraints, flexible adaptation)

Biological Resiliency Principles

Michael Levin's research reveals how biological systems achieve remarkable robustness through:

1. Anatomical Homeostasis

Biological systems actively maintain target morphology by continuously reducing error. Example: salamander limb regeneration stops when (and only when) a "correct salamander limb" is complete.

Applied to Skills: Skills should define clear success criteria and include mechanisms to validate progress toward those goals.

2. Error Correction

Systems activate effectors at multiple scales to reduce error relative to target states. Cells coordinate to repair damage and restore function.

Applied to Skills: Skills should include validation checkpoints that detect deviations from expected states and trigger corrective actions.

3. Robustness Through Collective Intelligence

Cell swarms exhibit both robustness (reliable outcomes) and plasticity (multiple valid paths). The system is resilient to perturbations while exploring solution space.

Applied to Skills: Skills should define required outcomes while allowing flexibility in implementation paths. Multiple approaches can achieve the same validated result.

4. Goal-Directed Behavior

Biological systems work toward specific outcomes across timescales, using feedback loops to navigate toward target states despite noise and variation.

Applied to Skills: Skills should specify goals explicitly and provide feedback mechanisms to assess progress and adjust course.

Resiliency Design Framework

Step 1: Assess Determinism Requirements

Classify your skill's determinism level:

High Determinism Indicators:

• Executes scripts or automated processes
• Modifies production systems or data
• Requires specific file formats or structures
• Depends on external tool outputs
• Has compliance or security requirements
• Must produce bit-perfect reproducible results

Medium Determinism Indicators:

• Follows structured workflows with clear phases
• Creates artifacts with defined schemas
• Interacts with version control or APIs
• Requires validation of intermediate states
• Has quality gates or checkpoints

Low Determinism Indicators:

• Explores problem space creatively
• Generates ideas or content
• Provides guidance without strict structure
• Adapts to user preferences dynamically
• Has subjective success criteria

Step 2: Design Resiliency Mechanisms

Based on determinism level, implement appropriate resiliency patterns:

High Determinism Skills: Comprehensive Resiliency

Pre-Execution Validation:

• Validate all inputs against schemas
• Check preconditions (file existence, permissions, dependencies)
• Verify tool availability and versions
• Create backup/snapshot points
• Document expected state

Execution Monitoring:

• Validate each step's output before proceeding
• Log state transitions and decisions
• Implement retry mechanisms with exponential backoff
• Set timeouts for external operations
• Capture error context for debugging

Post-Execution Verification:

• Verify final output against success criteria
• Run test suites or validation scripts
• Check system state consistency
• Generate execution report
• Provide rollback instructions if validation fails

Recovery Procedures:

• Include explicit rollback steps
• Document manual recovery procedures
• Provide diagnostic scripts
• Log all state changes for audit trail

Example Structure:

high-determinism-skill/
├── SKILL.md                    # Core workflow with validation checkpoints
├── scripts/
│   ├── validate-input.py      # Pre-execution validation
│   ├── verify-output.py       # Post-execution verification
│   └── rollback.sh            # Recovery procedure
├── references/
│   ├── schemas.md             # Input/output schemas
│   └── recovery-guide.md      # Detailed recovery procedures
└── examples/
    └── validated-workflow.md  # Example with all checkpoints

Medium Determinism Skills: Balanced Resiliency

Checkpoint Validation:

• Define phase boundaries with clear exit criteria
• Validate artifacts at checkpoints
• Provide status summaries
• Enable phase restart without full restart

Error Detection:

• Identify common failure modes
• Provide clear error messages with context
• Suggest corrective actions
• Document troubleshooting steps

State Management:

• Create summary files tracking progress
• Enable resumption from checkpoints
• Document current state clearly
• Track key decisions

Example Structure:

medium-determinism-skill/
├── SKILL.md                        # Workflow with checkpoints
├── scripts/
│   └── validate-checkpoint.sh     # Phase validation
├── references/
│   └── troubleshooting.md         # Common issues and fixes
└── examples/
    └── checkpoint-workflow.md     # Example with phase validation

Low Determinism Skills: Minimal Resiliency

Guidance and Constraints:

• Provide principles, not procedures
• Suggest rather than prescribe
• Enable exploration within bounds
• Document anti-patterns to avoid

Adaptive Feedback:

• Ask for user feedback on direction
• Offer multiple valid approaches
• Adjust based on user preferences
• Validate satisfaction, not output

Example Structure:

low-determinism-skill/
├── SKILL.md                    # Principles and guidance
└── references/
    └── examples.md            # Diverse example approaches

Step 3: Implement Feedback Loops

Design homeostatic feedback mechanisms appropriate to determinism level:

For High Determinism:

• Automated validation at each step
• Fail-fast on validation errors
• Required checkpoints before proceeding
• Automated rollback on critical failures

For Medium Determinism:

• Explicit user confirmation at phase boundaries
• Validation with override capability
• Warning system for potential issues
• Manual recovery guidance

For Low Determinism:

• Periodic progress checks
• User satisfaction validation
• Flexible adaptation to feedback
• Optional validation only

Step 4: Document Target States

Clearly define success criteria and target states:

For Scripts and Automation:

• Exact file structures, formats, schemas
• Required tool versions and configurations
• Expected output patterns
• Performance criteria (timeouts, resource limits)

For Workflows:

• Deliverables for each phase
• Quality criteria for artifacts
• Integration points and dependencies
• Completion criteria

For Guidance Skills:

• Principles to follow
• Anti-patterns to avoid
• Subjective quality indicators
• User satisfaction metrics

Implementation Patterns

Pattern 1: The Validation Checkpoint

Insert validation checkpoints at critical boundaries:

## Phase 1: Initialize

[Instructions for phase 1]

**Validation Checkpoint:**
Before proceeding to Phase 2, verify:
- [ ] Input file exists and is readable
- [ ] Required tools are available (version >= X.Y)
- [ ] Output directory has write permissions
- [ ] No conflicting processes running

Run `scripts/validate-phase1.sh` to check all conditions.

If validation fails, see `references/troubleshooting.md` Section 1.

Pattern 2: The Self-Correcting Loop

Implement error detection and correction:

## Step 3: Process Data

Process the input file using the following approach:

1. Load and validate input schema
2. Transform data according to rules
3. **Validation:** Check output against schema
4. **If validation fails:**
   - Log specific validation errors
   - Attempt automatic correction (see `scripts/auto-correct.py`)
   - If auto-correction fails, provide manual fix guidance
   - Retry validation after fixes
5. Proceed only after validation passes

Pattern 3: The Resilient Script

For skills that bundle scripts, add resiliency directly:

#!/usr/bin/env python3
"""
Resilient data processor with built-in validation and error handling.
"""

def validate_input(data):
    """Pre-execution validation."""
    # Check schema, types, ranges, etc.
    if not meets_criteria(data):
        raise ValueError("Input validation failed: [specific error]")

def process_with_retry(data, max_retries=3):
    """Process with exponential backoff retry."""
    for attempt in range(max_retries):
        try:
            result = process(data)
            if validate_output(result):
                return result
            else:
                log_warning(f"Output validation failed on attempt {attempt + 1}")
        except Exception as e:
            if attempt == max_retries - 1:
                raise
            wait_time = 2 ** attempt
            log_info(f"Retry in {wait_time}s after error: {e}")
            time.sleep(wait_time)

def validate_output(result):
    """Post-execution validation."""
    # Verify result meets success criteria
    return meets_success_criteria(result)

Pattern 4: The Homeostatic Summary

Maintain state awareness through summary files:

## summary.md Template

---
status: in-progress  # or: completed, blocked, failed
phase: execution
determinism: high    # high, medium, low
last-validated: 2026-01-07T15:30:00Z
---

# Execution Summary

## Current State
Currently executing Phase 2 of 4. Last checkpoint passed at 15:25.

## Validations Passed
- ✅ Input validation (15:20)
- ✅ Phase 1 completion (15:25)
- ⏳ Phase 2 in progress

## Known Issues
None

## Next Checkpoint
Phase 2 validation at step 5

## Rollback Point
Can rollback to Phase 1 completion using `scripts/rollback.sh phase1`

Resiliency Checklist by Determinism Level

High Determinism Checklist

• Input validation with schema checking
• Precondition verification (files, tools, permissions)
• Step-by-step output validation
• Retry mechanisms for transient failures
• Timeout handling for external operations
• Post-execution verification
• Rollback procedures documented
• Error logging with full context
• Recovery scripts provided
• Test suite for validation logic

Medium Determinism Checklist

• Phase boundaries with clear exit criteria
• Checkpoint validation
• Progress tracking in summary file
• Common error detection
• Troubleshooting guide
• Restart-from-checkpoint capability
• User confirmation at key transitions
• Warning system for potential issues

Low Determinism Checklist

• Clear principles documented
• Anti-patterns identified
• Multiple example approaches
• Periodic progress checks
• User satisfaction validation
• Flexible adaptation to feedback

Anti-Patterns to Avoid

Over-Engineering Low Determinism Skills

Wrong: Adding extensive validation to creative or exploratory skills. Right: Provide principles and guidance, allow flexibility.

Under-Engineering High Determinism Skills

Wrong: Assuming scripts will work without validation. Right: Validate inputs, outputs, and intermediate states comprehensively.

Silent Failures

Wrong: Continuing execution after validation failures without alerting. Right: Fail fast and provide clear error context.

Validation Without Recovery

Wrong: Detecting errors but providing no path forward. Right: Include recovery procedures or rollback instructions for all validation failures.

False Homeostasis

Wrong: Checking that "something happened" without verifying it's correct. Right: Validate against explicit success criteria, not just completion.

Testing Resiliency

Validate your skill's resiliency by testing:

Failure Modes:

• Missing files or dependencies
• Invalid inputs
• Interrupted execution
• Tool failures or timeouts
• Partial completion scenarios

Recovery Paths:

• Rollback procedures work correctly
• Checkpoints enable restart
• Error messages are actionable
• Recovery scripts function as documented

Validation Effectiveness:

• False positives (flagging valid states as errors)
• False negatives (missing actual errors)
• Validation coverage (all critical states checked)

Additional Resources

Reference Files

For deeper understanding of resiliency concepts:

• references/levin-principles.md - Michael Levin's biological resiliency principles in detail
• references/determinism-assessment.md - Comprehensive guide to assessing skill determinism
• references/validation-patterns.md - Extensive validation pattern library

Example Files

Working examples demonstrating resiliency levels:

• examples/high-determinism-deployment.md - Deployment skill with full resiliency
• examples/medium-determinism-workflow.md - Workflow skill with checkpoints
• examples/low-determinism-creative.md - Creative skill with minimal constraints

Scripts

Utility scripts for implementing resiliency:

• scripts/generate-validation.py - Generate validation scripts from schemas
• scripts/test-resiliency.sh - Test skill under failure conditions

Remember: Resiliency is not about preventing all failures—it's about detecting deviations from target states and providing pathways to correction. Match resiliency mechanisms to determinism requirements: over-engineering adds friction, under-engineering invites failure.