Stacks / 7 workflow stages · v3.1.0 · reviewed 2026-09-05
Evaluate and Harden an AI System
Turn subjective model quality into repeatable tests, traces, adversarial checks, and a release gate.
Best for
Teams moving an AI feature from 'it seems to work' toward measurable reliability and safer production behavior.
Outcome
A practical evaluation loop covering task quality, regressions, tracing, retrieval quality where relevant, and adversarial/security testing.
Not for
Early ideation where the product task itself is still undefined and there is nothing stable enough to evaluate.
Operator guidance
Best whenThe AI system is moving toward production or changes frequently enough to create regression risk.
Avoid whenThere is no representative task set, owner, or release decision tied to the scores.
CompatibilityUse shared case IDs across traces, graders, red-team runs, and release gates; separate model quality from system and tool failures.
Use this as a decision framework, not a mandatory shopping list. Swap or skip layers based on your existing stack, constraints, and risk profile.
1. Define the eval set
Start by turning real product behavior and failures into a repeatable evaluation loop.
- Hamel Husain: It's Hard to Eval Is a Product Smell
Recommendation details
Why: Consider Hamel Husain: It's Hard to Eval Is a Product Smell as an alternative for 1. Define the eval set; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Product-oriented argument for making AI outputs inherently easy to verify before layering on evaluation tooling.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: Hamel Husain: It's Hard to Eval Is a Product Smell official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
Recommendation details
Best when: Concrete example of improving prompts through an evaluation flywheel.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: OpenAI Evaluation Flywheel official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
2. Run regression tests
Track task-level quality across prompts, models, and application changes instead of comparing outputs by feel.
- Promptfoo
Recommendation details
Why: Consider Promptfoo as an alternative for 2. Run regression tests; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Open-source LLM eval and red-team framework.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: Promptfoo official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
- DeepEval
Recommendation details
Why: Consider DeepEval as an alternative for 2. Run regression tests; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Code-first evaluation framework.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: DeepEval official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
Recommendation details
Best when: Evals, traces and production quality workflows.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: Braintrust official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
3. Evaluate retrieval if applicable
Measure whether retrieval systems actually return relevant evidence and grounded answers.
- DeepEval
Recommendation details
Why: Consider DeepEval as an alternative for 3. Evaluate retrieval if applicable; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Code-first evaluation framework.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: DeepEval official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
Recommendation details
Best when: RAG-focused evaluation framework.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: Ragas official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
4. Trace failures
Inspect traces and model behavior so a failed evaluation can be diagnosed rather than merely counted.
- Langfuse
Recommendation details
Why: Consider Langfuse as an alternative for 4. Trace failures; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Open-source LLM observability and evals.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: Langfuse official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
- LangSmith
Recommendation details
Why: Consider LangSmith as an alternative for 4. Trace failures; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Tracing and observability platform for agent and LLM applications.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: LangSmith official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
Recommendation details
Best when: Open-source tracing and evaluation.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: Arize Phoenix official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
5. Red-team the application
Probe common LLM vulnerabilities and failure modes with a purpose-built red-team tool.
- garak Website
Recommendation details
Why: Consider garak Website as an alternative for 5. Red-team the application; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Project site for garak, the LLM vulnerability scanner.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: garak Website official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
Recommendation details
Best when: LLM vulnerability scanner.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: NVIDIA garak official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
6. Threat-model systematically
Use a structured adversarial framework to reason about attacks beyond individual prompt-injection examples.
- OWASP GenAI / LLM Top 10 — 2026
Recommendation details
Why: Consider OWASP GenAI / LLM Top 10 — 2026 as an alternative for 6. Threat-model systematically; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Current OWASP 2026 risk list for GenAI and LLM applications.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: OWASP GenAI / LLM Top 10 — 2026 official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
Recommendation details
Best when: Adversarial threat knowledge base for AI.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: MITRE ATLAS official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
7. Make release criteria explicit
Pair quality metrics with concrete security and deployment considerations before exposing the system more broadly.
- OWASP GenAI Security Project
Recommendation details
Why: Consider OWASP GenAI Security Project as an alternative for 7. Make release criteria explicit; compare its documented capabilities and operating constraints with the primary recommendation.
Best when: Core security guidance for GenAI systems.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: OWASP GenAI Security Project official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
Recommendation details
Best when: Anthropic plugin with edit warnings and diff review for common vulnerability classes.
Avoid when: Avoid when the current official source does not support your required deployment, data-handling, or integration constraints.
Compatibility: Validate interfaces, data formats, and operational handoffs with the adjacent stack stages.
Prerequisites:
- Review the current official source and implementation requirements.
Evidence: Anthropic Security Guidance official source (2026-09-05)
Verified 2026-09-05 · Medium confidence
Implementation references
-
OpenAI Evals Cookbook ↗
Implementation examples for datasets, graders, trace evaluation, and iterative improvement loops.
developers.openai.com · cookbook -
Anthropic: Demystifying Evals for AI Agents ↗
A practical framework for matching evaluation methods to long-horizon, tool-using agent behavior.
anthropic.com · reference architecture -
Eugene Yan: Fixing the Evaluation Process ↗
A concise process for turning observed failures into labeled cases, hypotheses, experiments, and release decisions.
eugeneyan.com · reference architecture
Need this workflow built and running, not just linked? EE Solutions implements agentic stacks like this one. EE Solutions is a senior technology team for private capital firms and their portfolio companies.
Talk to EE Solutions ↗