Agentic eval

Agentic Evaluation Patterns

Patterns for self-improvement through iterative evaluation and refinement.

Overview

Evaluation patterns enable agents to assess and improve their own outputs, moving beyond single-shot generation to iterative refinement loops.

Generate → Evaluate → Critique → Refine → Output
    ↑                              │
    └──────────────────────────────┘

When to Use

• Quality-critical generation: Code, reports, analysis requiring high accuracy
• Tasks with clear evaluation criteria: Defined success metrics exist
• Content requiring specific standards: Style guides, compliance, formatting
• Iterative improvement workflows: Drafts, plans, prompts, or code need critique and refinement before delivery

Do Not Use This Skill For

• Formal compliance claims, certification, or audit sign-off without approved organisational evidence
• Production monitoring design, runtime governance, access control, or policy enforcement
• Broad statistical evaluation methodology, benchmark design, or model selection studies
• Security, privacy, or safety review where evaluation loops are not the central concern

---

Expected Outputs

When applying this skill, produce the assets needed to run and inspect an evaluation loop:

• evaluation criteria or rubric
• generator, evaluator, and optimiser responsibilities
• iteration limit and convergence policy
• structured evaluator output format
• failure handling for malformed evaluations or non-improving outputs
• logging fields for debugging and review
• final recommendation or refined output with evaluation history summary

Treat code blocks in this skill as adaptable Python-style skeletons. Replace llm, run_tests, model settings, and logging functions with project-specific implementations.

Evaluation History Schema

Log enough detail to explain why the loop stopped and which candidate was selected:

{
    "iteration": 1,
    "candidate_id": "draft-1",
    "overall_score": 0.82,
    "dimension_scores": {"accuracy": 0.8, "clarity": 0.9},
    "feedback": ["Add evidence for the risk rating."],
    "parse_status": "valid",
    "improvement_delta": 0.07,
    "stopping_reason": null
}

Use parse_status values such as valid, retry_valid, or failed_closed. Final summaries should include the best candidate, the stopping reason, and any evaluator uncertainty or human-review escalation.

Pattern 1: Basic Reflection

Agent evaluates and improves its own output through self-critique.

import json


def parse_evaluation(raw_response: str, schema_prompt: str) -> dict:
    try:
        return json.loads(raw_response)
    except json.JSONDecodeError:
        retry = llm(f"Return only valid JSON for this schema: {schema_prompt}\nResponse: {raw_response}")
        try:
            return json.loads(retry)
        except json.JSONDecodeError as exc:
            raise ValueError({"parse_status": "failed_closed", "raw_response": raw_response}) from exc


def reflect_and_refine(task: str, criteria: list[str], max_iterations: int = 3) -> str:
    """Generate with reflection loop."""
    output = llm(f"Complete this task:\n{task}")
    schema_prompt = '{"criterion": {"status": "PASS|FAIL", "feedback": "..."}}'
    
    for i in range(max_iterations):
        # Self-critique
        critique = llm(f"""
        Evaluate this output against criteria: {criteria}
        Output: {output}
        Rate each: PASS/FAIL with feedback as JSON.
        """)
        
        critique_data = parse_evaluation(critique, schema_prompt)
        all_pass = all(c["status"] == "PASS" for c in critique_data.values())
        if all_pass:
            return output
        
        # Refine based on critique
        failed = {k: v["feedback"] for k, v in critique_data.items() if v["status"] == "FAIL"}
        output = llm(f"Improve to address: {failed}\nOriginal: {output}")
    
    return output

Key insight: Use structured JSON output for reliable parsing of critique results, and handle parsing failures as evaluation failures rather than silently accepting the draft.

---

Pattern 2: Evaluator-Optimizer

Separate generation and evaluation into distinct components for clearer responsibilities.

import json


class EvaluatorOptimizer:
    def __init__(self, score_threshold: float = 0.8):
        self.score_threshold = score_threshold
    
    def generate(self, task: str) -> str:
        return llm(f"Complete: {task}")

    def parse_evaluation(self, raw_response: str) -> dict:
        schema_prompt = '{"overall_score": 0.0, "dimensions": {"accuracy": 0.0, "clarity": 0.0}, "evidence": []}'
        try:
            return json.loads(raw_response)
        except json.JSONDecodeError:
            retry = llm(f"Return only valid JSON for this schema: {schema_prompt}\nResponse: {raw_response}")
            try:
                return json.loads(retry)
            except json.JSONDecodeError as exc:
                raise ValueError({"parse_status": "failed_closed", "raw_response": raw_response}) from exc
    
    def evaluate(self, output: str, task: str) -> dict:
        raw_evaluation = llm(f"""
        Evaluate output for task: {task}
        Output: {output}
        Return JSON: {{"overall_score": 0-1, "dimensions": {{"accuracy": ..., "clarity": ...}}, "evidence": [...]}}
        """)
        return self.parse_evaluation(raw_evaluation)
    
    def optimize(self, output: str, feedback: dict) -> str:
        return llm(f"Improve based on feedback: {feedback}\nOutput: {output}")
    
    def run(self, task: str, max_iterations: int = 3) -> str:
        output = self.generate(task)
        previous_score = 0.0
        for _ in range(max_iterations):
            evaluation = self.evaluate(output, task)
            current_score = evaluation["overall_score"]
            if current_score >= self.score_threshold:
                break
            if current_score <= previous_score:
                break
            previous_score = current_score
            output = self.optimize(output, evaluation)
        return output

---

Pattern 3: Code-Specific Reflection

Test-driven refinement loop for code generation.

class CodeReflector:
    def reflect_and_fix(self, spec: str, max_iterations: int = 3) -> str:
        code = llm(f"Write Python code for: {spec}")
        tests = llm(f"Generate pytest tests for: {spec}\nCode: {code}")
        
        for _ in range(max_iterations):
            result = run_tests(code, tests)
            if result["success"]:
                return code
            code = llm(f"Fix error: {result['error']}\nCode: {code}")
        return code

---

Evaluator Reliability

Evaluation loops are only useful when the evaluator is harder to fool than the generator. For high-risk or quality-critical work:

• Anchor the evaluator with a concrete rubric and examples of passing and failing outputs.
• Ask for evidence, failure reasons, and uncertainty, not just a score.
• Randomise or blind comparison order when comparing alternatives to reduce position bias.
• Repeat judging or use multiple evaluators when scores are close or the decision is consequential.
• Escalate to human review when evaluator outputs conflict, confidence is low, or policy-sensitive content is involved.
• Track whether scores improve over iterations; stop when improvement stalls to avoid overfitting to the evaluator.

---

Evaluation Strategies

Outcome-Based

Evaluate whether output achieves the expected result.

def evaluate_outcome(task: str, output: str, expected: str) -> str:
    return llm(f"Does output achieve expected outcome? Task: {task}, Expected: {expected}, Output: {output}")

LLM-as-Judge

Use LLM to compare and rank outputs.

def llm_judge(output_a: str, output_b: str, criteria: str) -> str:
    return llm(f"Compare outputs A and B for {criteria}. Which is better and why?")

Rubric-Based

Score outputs against weighted dimensions.

RUBRIC = {
    "accuracy": {"weight": 0.4},
    "clarity": {"weight": 0.3},
    "completeness": {"weight": 0.3}
}

def evaluate_with_rubric(output: str, rubric: dict) -> float:
    raw_scores = llm(f"Rate 1-5 for each dimension as JSON: {list(rubric.keys())}\nOutput: {output}")
    try:
        scores = json.loads(raw_scores)
    except json.JSONDecodeError:
        retry = llm(f"Return only valid JSON scores for: {list(rubric.keys())}\nResponse: {raw_scores}")
        scores = json.loads(retry)
    return sum(scores[d] * rubric[d]["weight"] for d in rubric) / 5

---

Best Practices

Practice	Rationale
Clear criteria	Define specific, measurable evaluation criteria upfront
Iteration limits	Set max iterations (3-5) to prevent infinite loops
Convergence check	Stop if output score isn't improving between iterations
Log history	Keep full trajectory for debugging and analysis
Structured output	Use JSON for reliable parsing of evaluation results
Evaluator calibration	Use examples, rubrics, and evidence requirements to reduce judge drift
Human escalation	Route low-confidence or high-impact decisions to a reviewer

---

Failure Handling

Plan for evaluation failures before running the loop:

• If evaluator JSON is malformed, retry once with the schema and then fail closed with the raw response logged.
• If the evaluator gives no evidence, treat the score as unreliable and request evidence before refining.
• If output quality stops improving, return the best-scored candidate and include the stopping reason.
• If evaluation criteria conflict, pause and ask for priority order rather than optimising against incompatible goals.
• If generated tests or checks are themselves suspect, review or regenerate the tests before trusting the score.

---

Quick Start Checklist

## Evaluation Implementation Checklist

### Setup
- [ ] Define evaluation criteria/rubric
- [ ] Set score threshold for "good enough"
- [ ] Configure max iterations (default: 3)

### Implementation
- [ ] Implement generate() function
- [ ] Implement evaluate() function with structured output
- [ ] Implement optimize() function
- [ ] Wire up the refinement loop

### Safety
- [ ] Add convergence detection
- [ ] Log all iterations for debugging
- [ ] Handle evaluation parse failures gracefully
- [ ] Escalate low-confidence or high-impact decisions to human review