The right tool for the right job — data foundations, when to reach for an agent versus a statistical test, and how to build reproducible, optimizable LLM pipelines with DSPy.
Takeaway 1. An agent can only navigate the chemical domain its training data represents — homogeneous, dense coverage supports interpolation and extrapolation, while sparse or heterogeneous data (and uncaught characterization artifacts like bad calibration or matrix effects) silently breaks any downstream model.
Takeaway 2. An agent is an LLM in a harness that calls tools repeatedly in a loop, deciding its next step from the last result — ReAct formalizes this by interleaving chain-of-thought reasoning with tool calls, and Coscientist (built on Gemini 2.0) applies the same idea as a multi-agent system for generating scientific hypotheses.
Takeaway 3. Not every problem needs an LLM "brain" — use a rubric (can it be represented mathematically? does a statistical test already exist? is there an off-the-shelf model?) to route A/B tests, regressions, and classification to conventional tools, reserving agents for tasks like open-ended semantic analysis.
Takeaway 4. Hand-tuned prompts are not reproducible and silently degrade when a model provider ships an update — DSPy compiles a program from a typed signature, a module, a metric, and an optimizer instead, so any one part can be swapped (a cheaper model, a new metric) without rewriting the rest.
Takeaway 5. GEPA's reflection-based optimization beat GRPO reinforcement learning with 35× fewer rollouts and outperformed MIPROv2 by 10%, and Shopify used it to move a GPT-5 task onto a small Qwen model at ~75× lower cost — evidence that optimized small models can match hand-prompted frontier models.
Takeaway 6. Self-driving labs need more than a capable model — physical stops, human-in-the-loop checkpoints, and auditable (git-tracked, annotated) decision logs separate credible autonomous pipelines from irreproducible ones.