Europe 2026
The maturity phases of running evals — Phil Hetzel, Braintrust
Overview
This talk outlines the maturity phases of running evaluations for AI agents, emphasizing that evals are crucial for ensuring agent quality, mitigating risks, and understanding performance improvements. The speaker suggests a progression from initial human-based assessments to more automated and complex evaluation strategies as agent complexity increases. The core idea is to systematically build confidence in agent behavior before and after deployment.
Who should watch
- AI Engineers
- Product Managers
- Builders developing AI agents
- Anyone responsible for agent quality and risk mitigation
- Teams looking to move AI agents from proof-of-concept to production
Key takeaways
- Evals are essential for agent quality, serving as a defense against reputational, system, and compliance risks, and as an offense to measure improvements.
- Unlike unit tests, evals should focus on specific failure modes rather than exhaustive testing to avoid hindering development velocity.
- Initial evaluation can start with human annotation, documenting thumbs up/down decisions and justifications to extract domain knowledge.
- As agents become more complex, especially with tool or MCP calls, evaluating entire traces and representing external system states becomes critical.
- LLMs can be used as judges for other LLMs, but these judges themselves require evaluation to ensure alignment with human judgment.
- Capturing production traces and rerunning them through evals in an offline environment helps guide agent improvements.
- Deterministic checks, such as limiting tool calls or token usage, can complement LLM-based evaluations.
- Representing external system state within traces or using timestamped queries can help approximate production environments during evals.
Notable quotes
*Evals are a both a defense against those types of risks. But they're also they can play offense with evals in knowing with each tweak that you make to your agent how it's improving.*
*Eval results don't need to be perfect. Sometimes they can be directional.*
*The more complex agent you're that you're building the more vectors there are for failure the more failure modes you that you may need to account for.*
Unofficial community note. Prefer the recording for nuance.