Evals became the bottleneck. Teams that shipped fast had evals running before features, not after.
Model evals vs product evals. Benchmarks tell you about the model. Production metrics tell you about your system.
Agent evaluation got structured. Four buckets emerged for evaluating agentic systems systematically.
Monitoring closed the loop. Real-time signals feeding back into development. The flywheel that compounds quality.
Quality Verification
Quality verification became the bottleneck for shipping AI.
Evals
Model Evals vs Product Evals
Model evals test the underlying model (benchmarks, capabilities). Product evals test your system (task completion, user success). Both matter—but product evals determine if you ship.
Evals
Evals
Offline Quality Assurance
Run before deployment. Test against known datasets. Catch regressions before users see them. Answer: "Is this change safe to ship?"
→
Monitoring
Online Quality Assurance
Run in production. Track real user interactions. Catch issues evals missed. Answer: "Is this working for actual users?"
Evals
Evals vs Monitoring: When They Run
Evals run offline before deployment. Monitoring runs online in production. Both essential—different timing, different purpose.
Evals
The eval mental model: What are you actually testing?
Every eval answers one of three questions. (1) Can it do the task at all? Capability. (2) Does it still do the task after changes? Regression. (3) Does it do the task the way we want? Alignment. Know which question you're asking.
Evals
Four Buckets for Agent Evaluation
1. Task Completion
Did the agent achieve the goal? Binary success/failure on well-defined objectives. The baseline metric.
2. Trajectory Quality
How did it get there? Efficient tool use, sensible step ordering, recovery from errors. The path matters.
3. Safety & Boundaries
Did it stay in bounds? No unauthorized actions, proper escalation, respecting guardrails. Trust requires limits.
4. Resource Efficiency
What did it cost? Tokens consumed, API calls made, time elapsed. Efficiency at scale.
Evals
The Grader Stack: Who Evaluates?
Deterministic checks first (fast, cheap). LLM-as-judge for scale (the 2025 breakthrough). Human review for calibration and edge cases.
Evals
Capability Evals
"Can it do new things?"
Testing new features. Expanding to new domains. Pushing boundaries. Run when adding capabilities.
↔
Regression Evals
"Does it still work?"
Catching breakage. Model updates, prompt changes, dependency shifts. Run on every change. Non-negotiable.
Monitoring
The Eval Flywheel: How Quality Compounds
Ship → Observe → Curate failures into eval cases → Eval before next deploy → Improve → Ship again. Each cycle makes the system more robust.
Building AI Products
If You're Building AI Products, Know This
Evals before features. You can't iterate fast without fast feedback. Build the eval harness first.
LLM-as-judge scales. Human review doesn't. Calibrate your LLM graders against human judgment, then trust them.
Regression tests are sacred. Every production failure becomes a test case. The suite only grows.
Monitor implicit signals. Users don't file bug reports. They regenerate, abandon, or leave. Watch behavior.
Close the loop. Production → evals → improvements → production. The flywheel compounds.
Output Layer: Key Takeaways
Evals are the shipping bottleneck. Fast eval cycles = fast iteration. Invest here first.
Model evals ≠ product evals. Benchmarks don't tell you if users succeed. Test what matters.
LLM-as-judge changed evaluation. Scale beyond human capacity. Calibrate carefully.
The flywheel wins. Production failures → eval cases → prevented failures. Compound quality over time.
