Agent Evaluation

Agent evaluation measures whether an AI agent produces output that meets a defined quality bar. Without evaluation, you're guessing whether the agent works. With evaluation, you know: the agent passes 94% of accuracy checks, violates zero banned-phrase rules, and produces output rated 4.2/5 by human reviewers. Evaluation turns "it seems fine" into a measurable quality system.

The principle: define what "good" means before building the agent. Write the eval before the prompt. If you can't describe what a good output looks like in measurable terms, you can't evaluate whether the agent produces it.

The Evaluation Framework

4 dimensions of agent quality

Dimension priority

Accuracy > Compliance > Completeness > Quality. In that order.

• An accurate output that's slightly too long (compliance miss) is fixable
• A compliant output that contains fabricated data (accuracy miss) is dangerous
• An output that's missing one field (completeness miss) is incomplete but safe
• A low-quality output that's accurate and compliant just needs prompt improvement

Building Evals for GTM Agents

Eval structure

Every agent eval has three components:

1. TEST SET: A curated set of inputs with known-good outputs
2. EVAL CRITERIA: Specific, measurable checks applied to each output
3. SCORING: How to aggregate individual checks into an overall score

Creating the test set

Test set rules:

• Build the golden set from real data. Export 20-50 real prospect records. Have a human write the ideal output for each. This is the ground truth the agent is measured against
• Include bad data in the test set. Prospects with missing fields, incomplete LinkedIn profiles, companies with no funding data. The agent must handle bad input gracefully
• Add every failure to the regression set. When the agent produces a bad output in production, add that input + the correct output to the regression set. The test set grows stronger over time
• 20 examples minimum for any meaningful eval. Below 20, individual outliers dominate the score. Above 50, you get diminishing returns (unless the agent handles very diverse tasks)

Eval Criteria by Agent Type

Research Agent eval

Email Writer Agent eval

Reply Classifier Agent eval

Lead Scorer Agent eval

Automated vs Human Evaluation

When to use each

Eval layer rules

• Automated checks run on every output. Zero cost, instant feedback. Word count, banned phrases, format, schema. These are binary pass/fail. No reason to skip
• LLM-as-judge is a middle layer. Use a second LLM (can be the same model with a different prompt) to rate quality 1-5. Useful for flagging outputs that pass automated checks but feel robotic or irrelevant
• Human review is the calibration layer. Everything else is calibrated against human judgment. Without regular human review, automated checks may pass outputs that are technically compliant but qualitatively bad

The Eval Pipeline

How to run evals continuously

Agent produces output
  ↓
Layer 1: Automated checks (instant)
  - Format compliance (schema, word count, structure)
  - Rule compliance (banned phrases, em-dashes, subject line rules)
  - Accuracy cross-check (proper nouns in output exist in input)
  ↓
  PASS → Layer 2
  FAIL → Reject. Log failure. Regenerate or flag for human fix

Layer 2: LLM-as-judge (optional, seconds)
  - Quality rating 1-5
  - Naturalness check
  - Relevance to the prospect's situation
  ↓
  Score ≥ 3.5 → PASS → Output goes to production
  Score < 3.5 → Flag for human review

Layer 3: Human spot-check (daily/weekly)
  - Random sample of 10-20% of passed outputs
  - Human rates 1-5 on quality
  - Flags hallucinations, tone issues, irrelevant content
  - Results feed back into prompt improvement

Tracking Eval Results Over Time

The eval dashboard

Quality trend analysis

Eval-Driven Prompt Improvement

The improvement cycle

1. Run eval on current prompt (golden set + production sample)
2. Identify weakest criterion (lowest pass rate or score)
3. Analyze failures: what pattern causes the failures?
4. Modify the prompt to address the pattern
5. Re-run eval on the SAME test set
6. Compare results: did the change improve the weak criterion
   WITHOUT degrading other criteria?
7. If yes: deploy the new prompt
8. If no: revert and try a different approach

Improvement rules

• One prompt change per eval cycle. Changing 5 things at once makes it impossible to know which change helped. Change one instruction, one example, or one rule per iteration
• Always re-run the full golden set after a change. A change that fixes one problem may introduce another. The golden set catches regressions
• Track every prompt version with its eval results. Version 1: accuracy 91%, quality 3.8. Version 2: accuracy 94%, quality 4.1. Version 3: accuracy 93%, quality 4.3. The version history shows the improvement trajectory
• Don't optimize for the test set at the expense of production. If the prompt scores perfectly on the golden set but poorly on production samples, the golden set isn't representative. Add more diverse examples

Pre-Deploy Eval Checklist

Before deploying any agent or prompt change to production:

• [ ] Golden set (20+ examples) created with human-verified ideal outputs
• [ ] Edge case set (10+ examples) covering missing data, unusual inputs
• [ ] Automated checks implemented for all compliance rules
• [ ] Agent passes 95%+ on accuracy checks against golden set
• [ ] Agent passes 100% on compliance checks (word count, banned phrases, format)
• [ ] Agent passes 90%+ on completeness checks
• [ ] Human review of 20+ outputs scores ≥ 4.0/5 average quality
• [ ] Hallucination rate < 2% on golden set
• [ ] Regression set passes at 100% (no re-introduced old failures)
• [ ] Eval results documented with prompt version number

Anti-Pattern Check

• No eval before deployment. "The prompt looks good in testing" is not an eval. Build a golden set. Run automated checks. Score with human reviewers. Then deploy
• Eval set too small (5 examples). 5 examples means one failure changes the pass rate by 20%. Not meaningful. Minimum 20 examples for any eval
• No automated checks on production output. Automated checks are free and instant. Every output should pass word count, banned phrases, and format checks before being used. There's no reason to skip this
• Evaluating accuracy without ground truth. "The output seems correct" is not an accuracy check. Compare every factual claim in the output against the input data or a verified source. Ground truth or it's not an eval
• Golden set never updated. The golden set from 3 months ago doesn't include the new ICP segment, the updated messaging, or the recent edge cases. Update quarterly with fresh real-world examples
• Optimizing for test set at the expense of production. Agent passes 100% on the golden set but produces mediocre output on real prospects. The golden set is too easy or not representative. Add harder, more diverse examples
• No regression testing. A prompt change fixes one problem and re-introduces two old ones. Without a regression set, old failures silently return. Add every failure to the regression set and re-run after every change
• LLM-as-judge without human calibration. The judge model rates everything 4.5/5. Humans rate the same outputs 3.2/5. The judge is miscalibrated. Calibrate the judge prompt by comparing its ratings to human ratings on 50+ examples