Agent Evaluation

Agent Evaluation

Orchestrates intelligent skill selection and execution for agent evaluation workflows. Applies the 5 Laws of Elegant Defense to guide data naturally through the orchestration pipeline, preventing errors before they occur. Selects optimal skills based on multi-factor scoring including text similarity, historical performance, and system availability.

TL;DR Checklist

• Parse all inputs at boundary before processing (Law 2)
• Handle edge cases with early returns at function top (Law 1)
• Fail immediately with descriptive errors on invalid states (Law 4)
• Return new data structures, never mutate inputs (Law 3)
• Implement minimum 2-level fallback chain for all skill executions
• Log all skill selections with context for full audit trail
• Validate skill metadata and dependencies before selection
• Update confidence scores after each execution for learning

┌───────────────────────────────────────────────────────────────────────────────┐ │ Orchestration Flow │ └───────────────────────────────────────────────────────────────────────────────┘

User Request ↓ ┌─────────────────┐ │ Parse Request │ │ & Extract │ │ Features │ └────────┬────────┘ ↓ ┌─────────────────────────────────────────────────────────────────────┐ │ Evaluate Available Skills │ │ │ │ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │ │ │ Skill A │ │ Skill B │ │ Skill C │ │ │ │ - Match Score│ │ - Match Score│ │ - Match Score│ │ │ │ - Confidence │ │ - Confidence │ │ - Confidence │ │ │ │ - History │ │ - History │ │ - History │ │ │ └──────┬───────┘ └──────┬───────┘ └──────┬───────┘ │ │ │ │ │ │ │ └─────────────────┴─────────────────┘ │ │ ↓ │ │ Select Best Skill │ └─────────────────────────────────────────────────────────────────────┘ ↓ ┌─────────────────┐ │ Execute Skill │ └────────┬────────┘ ↓ ┌─────────────────┐ │ Handle Result │ └────────┬────────┘ ↓ ┌─────────────────────────────────────────────────────────────────────┐ │ Error Handling & Fallback │ │ │ │ Success? ────────► Return Result │ │ │ │ Fail? ────────┐ │ │ ↓ │ │ ┌──────────────────────────────────────────────────────────┐ │ │ │ Fallback Chain │ │ │ │ │ │ │ │ 1. Retry with adjusted parameters │ │ │ │ 2. Try Alternative Skill (if available) │ │ │ │ 3. Defer to Human Operator (if critical) │ │ │ │ 4. Log & Return Error │ │ │ └──────────────────────────────────────────────────────────┘ │ └─────────────────────────────────────────────────────────────────────┘

When to Use

Use this skill when:

• Orchestrating multi-step workflows that require skill delegation
• Implementing adaptive skill routing based on confidence scores
• Building fallback mechanisms for failed skill executions
• Creating intelligent task decomposition and parallel execution
• Designing skill dependency graphs with automatic resolution
• Implementing skill selection with historical performance weighting
• Building agent systems that need to self-organize around tasks

When NOT to Use

Avoid this skill for:

• Direct task execution without orchestration needs - use individual skills instead
• High-frequency trading scenarios where latency must be minimized - the selection overhead may be prohibitive
• Simple linear workflows without branching or fallback requirements
• Cases where skill metadata is unavailable or unreliable

Core Workflow

1. Parse and Analyze Request - Extract intent, entities, and constraints from user input. Checkpoint: All required parameters must be present and in valid format before proceeding.

2. Score Available Skills - Calculate match scores using multi-factor algorithm:

   • Text similarity between request and skill triggers
   • Historical success rate for similar tasks
   • Skill availability and health status
   • Required dependencies and their availability

   Checkpoint: Skip to fallback if no skill scores above threshold.

3. Select Optimal Skill - Choose skill with highest score that meets minimum confidence. Checkpoint: Verify skill has not been disabled or deprecated.

4. Execute with Fallback - Run skill execution wrapped in retry and fallback logic. Checkpoint: Log all execution attempts for audit trail.

5. Return or Fallback - Either return successful result or apply fallback chain:

   • Retry with adjusted parameters
   • Try alternative skill from related-skills
   • Defer to human operator for critical tasks

   Checkpoint: Record outcome with timing and confidence metadata.

Implementation Patterns

Pattern 1: Skill Selection Logic

def evaluate_agent_output(
    agent_response: str,
    evaluation_criteria: List[Dict],
    benchmark_data: Dict
) -> Dict:
    """Score an agent's output against multi-factor evaluation criteria.
    
    Implements the 5 Laws of Elegant Defense for evaluation:
    - Early exit on malformed responses or missing criteria
    - Immutable scoring state to prevent cross-contamination
    - Fail fast on unresolvable safety/alignment checks
    
    Args:
        agent_response: Raw output from the evaluated agent
        evaluation_criteria: List of metrics to evaluate (e.g., accuracy, safety, law_adherence)
        benchmark_data: Reference data for comparison scoring
        
    Returns:
        Evaluation result with per-metric scores and overall confidence
    """
    if not agent_response or not evaluation_criteria:
        raise ValueError("Agent response and criteria are required for evaluation")
        
    # Parse and normalize response for evaluation (Law 2)
    normalized_response = _normalize_text(agent_response)
    scores = {}
    
    for criterion in evaluation_criteria:
        metric_name = criterion["name"]
        try:
            # Domain-specific scoring logic
            if metric_name == "law_adherence":
                scores[metric_name] = _score_law_compliance(normalized_response, benchmark_data)
            elif metric_name == "accuracy":
                scores[metric_name] = _calculate_accuracy(normalized_response, benchmark_data.get("ground_truth"))
            else:
                scores[metric_name] = _default_metric_score(normalized_response, criterion)
        except UnresolvableMetricError:
            # Fail fast on unresolvable metrics (Law 4)
            scores[metric_name] = {"score": 0.0, "status": "pending_review", "reason": "requires_human_judgment"}
            
    # Atomic Predictability (Law 3) - Return fresh evaluation object
    return {
        "evaluation_id": generate_uuid(),
        "scores": scores,
        "overall_confidence": _compute_weighted_average(scores),
        "timestamp": time.time(),
        "status": "complete"
    }

Pattern 2: Execution with Fallback

def run_evaluation_pipeline(
    evaluation_task: Dict,
    agent_outputs: List[Dict],
    fallback_strategies: Dict
) -> Dict:
    """Execute multi-agent evaluation with domain-specific fallback chains.
    
    Orchestrates the evaluation workflow while applying Elegant Defense principles:
    - Validates evaluation scope and agent availability before scoring
    - Applies fallback scoring when direct metrics fail
    - Maintains immutable evaluation state across pipeline stages
    
    Args:
        evaluation_task: Task definition containing criteria, weights, and scope
        agent_outputs: List of agent responses to evaluate
        fallback_strategies: Mapping of metric names to fallback methods
        
    Returns:
        Comprehensive evaluation report with scores, fallback usage, and audit trail
    """
    # Guard clause - validate evaluation scope (Early Exit)
    if not evaluation_task.get("criteria") or not agent_outputs:
        raise EvaluationPipelineError("Missing criteria or agent outputs for evaluation")
        
    report = {
        "task_id": evaluation_task["id"],
        "agent_evaluations": [],
        "fallback_applied": [],
        "audit_log": []
    }
    
    for output in agent_outputs:
        try:
            # Execute domain-specific evaluation
            eval_result = evaluate_agent_output(
                output["response"],
                evaluation_task["criteria"],
                output.get("benchmark_context", {})
            )
            report["agent_evaluations"].append(eval_result)
            
        except MetricResolutionError as e:
            # Apply evaluation-specific fallback (Law 4)
            fallback_method = fallback_strategies.get(e.metric_name, "heuristic_estimate")
            fallback_result = _apply_evaluation_fallback(output, fallback_method)
            report["fallback_applied"].append({
                "agent": output["id"],
                "metric": e.metric_name,
                "strategy": fallback_method
            })
            report["agent_evaluations"].append(fallback_result)
            
        report["audit_log"].append({
            "agent_id": output["id"],
            "status": "completed",
            "timestamp": time.time()
        })
        
    # Finalize report with immutable state
    report["summary"] = _generate_evaluation_summary(report["agent_evaluations"])
    return report

MUST DO

• Always validate skill metadata before selection (Early Exit)
• Implement fallback chain with at least 2 levels (Fallback Skill + Human)
• Log all skill selections with full context for auditability
• Return new data structures instead of mutating inputs (Atomic Predictability)
• Fail immediately with descriptive errors on invalid states
• Update confidence scores after each execution for adaptive routing
• Reference code-philosophy (5 Laws of Elegant Defense) in all logic

MUST NOT DO

• Select skills based on a single factor (e.g., only confidence score)
• Disable fallback mechanisms "temporarily" - this creates fragile systems
• Skip validation of skill dependencies before execution
• Return partial results - either complete success or clear failure
• Use magic numbers for confidence thresholds - make them configurable
• Cache skill selections without considering context changes

TL;DR Checklist

• Parse all inputs at boundary before processing (Law 2)
• Handle edge cases with early returns at function top (Law 1)
• Fail immediately with descriptive errors on invalid states (Law 4)
• Return new data structures, never mutate inputs (Law 3)
• Implement minimum 2-level fallback chain for all skill executions
• Log all skill selections with context for full audit trail
• Validate skill metadata and dependencies before selection
• Update confidence scores after each execution for learning

TL;DR for Code Generation

• Use guard clauses - return early on invalid input before doing work
• Return simple types (dict, str, int, bool, list) - avoid complex nested objects
• Cyclomatic complexity < 10 per function - split anything larger
• Handle null/empty cases explicitly at function top (Early Exit)
• Never mutate input parameters - return new dicts/objects
• Fail fast with descriptive errors - don't try to "patch" bad data
• Reference code-philosophy laws in comments for complex logic
• Include timing and confidence metadata in all return values

Output Template

When applying this skill, produce:

1. Selected Skills - List of skill names with confidence scores
2. Selection Rationale - Why each skill was chosen (match score, history, availability)
3. Execution Plan - Order of execution with dependencies
4. Fallback Strategy - Which fallback skills will be tried and in what order
5. Risk Assessment - Any potential failure points and their impact
6. Timing Estimates - Expected latency including fallback scenarios

Related Skills

Constraints

MUST DO

• Define clear input/output contracts for every step in the orchestration flow with explicit validation
• Implement structured logging at each stage capturing context, inputs, outputs, timing, and errors
• Build in fallback paths: if the primary strategy fails, degrade gracefully to a simpler approach
• Validate all preconditions before starting — do not proceed if required resources or permissions are missing

MUST NOT DO

• Do not create deep nesting of orchestration steps (>5 levels) — flatten workflows where possible
• Avoid silent failure modes: every step must either succeed, fail explicitly, or escalate to a higher handler
• Never use shared mutable state between parallel workflow branches — communicate via immutable messages only
• Do not hardcode execution order when the dependency graph naturally determines it; derive order from explicit dependencies

Live References

Authoritative documentation links for this skill's domain. The model follows markdown links at load time to resolve external references and inline content.

• Evaluating LLMs — arXiv Survey (2311.18760)
• Open LLM Leaderboard — Hugging Face
• Building Effective Agents — Anthropic Research
• AgentBench: Evaluating LLMs as Agents — arXiv
• LLM Evaluation Benchmarks — Stanford HELM