Does GenAI "belong" to data scientists? — Phil Hetzel, Braintrust

Key Concepts

• Agentic Development: The process of building autonomous or semi-autonomous AI agents capable of performing tasks.
• Agent Quality Platform: A framework focusing on Evals (experimentation/testing) and Observability (production monitoring).
• Context Engineering: The practice of refining prompts and input data to influence LLM behavior, as opposed to traditional model training.
• LLM-as-a-Judge: Using an LLM to evaluate the outputs of another agent or model.
• Human-in-the-loop (HITL): Integrating domain experts into the annotation and evaluation workflow to ensure agent relevance.

---

1. Organizational Approaches to Agentic Development

Phil Hetzel identifies two distinct organizational archetypes in the industry:

• Traditional Enterprise: Often delegates agent development to existing Data Science or Machine Learning (ML) platform teams. This is driven by the assumption that because "AI" is in the name, it naturally falls under the purview of those who manage predictive models.
• AI Natives: These organizations lack legacy silos. They utilize small, cross-functional teams where product engineers and AI engineers work together, maintaining close proximity to the end-user problem.

2. Traditional ML vs. Generative AI (GenAI)

The speaker highlights a fundamental shift in the development lifecycle:

• Traditional ML: Involves data pipelines, feature engineering, model training, and rigorous cross-validation. The goal is to build a model from scratch.
• GenAI/Agents: The underlying model (LLM) is already built by providers like OpenAI, Anthropic, or Mistral. The "value add" shifts from training to context engineering and prompting.
• Key Difference: In GenAI, the surface area for evaluation is much broader, requiring functional performance testing rather than just technical metrics like precision, recall, or F1 scores.

3. The Case for and Against Data Scientists Owning Agents

Arguments for Data Scientist Ownership:

• Governance: They possess the expertise to manage risks associated with neural networks.
• Rigorous Mindset: They are accustomed to formal testing processes and production deployment standards.
• Technical Depth: They are best suited for fine-tuning open-source models when specific use cases demand it.

Arguments for Non-Data Scientist Ownership (Product/Systems Engineers):

• API Proficiency: Product engineers are already skilled at integrating external APIs, which is the core of agentic architecture.
• Systems Complexity: Distributed agents (supervisor agents calling sub-agents across different infrastructure) present systems engineering challenges rather than statistical ones.
• Domain Proximity: Subject matter experts (SMEs) and Product Managers have the best understanding of the problem, making them ideal for prompt engineering and human annotation.

4. Methodologies and Frameworks

• The "Adult in the Room" Role: Data scientists should act as advisors who provide guardrails, explaining the probabilistic nature of LLMs (token prediction) to prevent over-reliance on the model's "intelligence."
• Evaluation Pipeline:
  1. Experimentation: Use Evals to gain confidence before production.
  2. Production: Use Observability to monitor real-world performance.
  3. Feedback Loop: Continuously feed production data back into the offline evaluation dataset to improve the system.
• Human Annotation: Non-technical domain experts should review agent traces to provide qualitative feedback, which is essential for refining agent behavior.

5. Notable Quotes

• "The answer is always in the middle." — Regarding the debate over who should own agentic development.
• "[LLMs] are just predicting token after token. It doesn't actually know anything really. It's just a bunch of stats problems at the end of the day." — Emphasizing the need for data scientists to provide technical guardrails.

6. Synthesis and Conclusion

The primary takeaway is that agentic development should not be siloed within a single department. While Data Scientists and ML Engineers provide critical value in governance, evaluation metrics, and fine-tuning, the most successful agentic applications are built by diverse, cross-functional teams.

The ideal framework involves:

1. Data Scientists: Managing guardrails, LLM-as-a-judge metrics, and fine-tuning.
2. Product/Systems Engineers: Handling API integrations and distributed infrastructure.
3. Domain Experts: Leading prompt engineering and human annotation to ensure the agent solves the intended business problem.

The focus should shift from "who owns the tool" to "how do we solve the problem," treating agents as a collaborative product rather than a purely technical model.