Expert context is no longer a one-time setup field. You can edit it while an experiment is running, and the new content will be used by the AI on the next iteration it generates. This article explains how that works, when to update, and how to use expert context to steer an optimization that is already in motion.
Two Roles Your Expert Context Can Play
As an experiment progresses, the same expert-context field ends up doing two different jobs. It helps to think about them separately when deciding what to edit:
Background and grounding. The stable, factual description of the problem — what is being optimized, what each variable does, what interactions you know about. This is the part that anchors the AI's reasoning and usually does not change during a run. Treat it as the slow-moving foundation.
Directional steering. Softer guidance about where you would like the optimization to focus next — encouraging exploration of a region the data is hinting at, asking for more exploitation around a current best, calling out a hypothesis you would like the AI to test. This is the part that is most likely to evolve as you see results.
In the product these live in the same text field — there is no formal split — but framing your edits in your head this way prevents accidental over-writing of the foundational context. A useful pattern is to keep the background paragraphs at the top of the field, and add a short "Current focus" section at the bottom that you revise iteration to iteration.
What Happens When You Edit Mid-Run
Already-generated recommendations don't change. Whatever the AI has already proposed stays exactly as it was — you'll see a banner in the UI reminding you of this. Edits take effect on the next batch of recommendations.
Both the initial-design and model-guided phases use it. If you are still in the initial-design phase, the updated context shifts where the next warm-start points are placed. If you are already in the model-guided phase, the updated context is fed into the AI's hypothesis generation alongside the surrogate model's suggestions.
The algorithm will not blindly follow you. Once data has been collected, the model weighs your context against what the experiments actually show. If your context contradicts the data, the data wins — the model will not follow expert assumptions that the observations have already refuted. This is intentional: it stops a confidently-written but wrong piece of context from derailing a converging optimization.
Insights are rewritten using the latest context. The natural-language insights attached to each iteration are produced by an LLM that is given the current expert context as part of its prompt. Editing the context immediately changes the framing, vocabulary, and reasoning of new insights — so updating mid-run is also a quick way to get explanations that better match how you think about the problem now.
When to Update — and When Not To
Good reasons to update:
You learned something the AI can't see yet. An off-platform measurement, a side observation from the lab, a hint from a colleague — anything that changes how you would describe the problem if a new collaborator joined today.
A hypothesis was tested. You wrote down a guess at the start ("I think pyrrolidine in DMF will dominate"). After a few iterations, the data has either confirmed it or pointed elsewhere. Update the context to reflect what the data actually shows, so the AI is not nudged toward the old guess on every subsequent iteration. This is the single most common reason intuitions become harmful — they get written once and never revisited.
You want to shift the exploration/exploitation balance. Late in a run, you may want to lock in around a current best and ask the AI to fine-tune; early on, you may want to push it toward an undersampled region. A short, directive sentence ("Focus the next iterations on the low-temperature, high-catalyst-loading corner") is enough.
A constraint or sensitivity became clearer. If, mid-run, you discover that conditions above some threshold are unsafe, unstable, or scientifically uninteresting, add that to the context immediately.
Reasons not to update:
The model is not proposing your favorite region. If the surrogate model has data and disagrees with your prior intuition, that is usually a feature, not a bug. Don't write the model out of disagreements you can't back up with data.
You want to tweak a single recommendation. Edits affect future iterations, not the current one. If a specific proposed point is unworkable, use the platform's tools for skipping or replacing recommendations rather than rewriting the context.
You're chasing noise. One surprising result is not yet a pattern. Wait until you see the trend repeat before encoding it in context.
A Note on Intuitions
Intuitions are welcome in expert context. The hunches you'd try first if you were running the experiment by hand are often genuinely informative, and the AI can usually do better with them than without. But there is one rule of thumb worth following:
If your intuition has been tested by a few iterations and the data didn't bear it out, edit it out of the context. Otherwise the AI keeps pulling on it iteration after iteration, even after the experiment has moved on.
A useful pattern is to phrase intuitions as hypotheses ("I expect higher catalyst loading to help — worth testing early") rather than as facts ("Higher catalyst loading is better"). Hypotheses are easy to update once they've been tested; facts feel awkward to revise.
Example Edits
Early in a run, encoding a hypothesis you want tested:
Current focus. Based on prior work, we suspect that the morpholine and pyrrolidine catalysts in moderately polar solvents will outperform the others. It is worth testing this hypothesis explicitly within the first few iterations before exploring more broadly.
Mid-run, after data has invalidated that hypothesis:
Current focus. Initial expectations around the morpholine/pyrrolidine catalysts in polar solvents have not been confirmed — yields with those combinations have plateaued in the 40–55% range. Recent results suggest the higher-yield region lies in the proline + lower-temperature regime. Prioritize exploring that area in the next iterations.
Late in a run, asking for refinement:
Current focus. The optimization has stabilized around a region with yields above 75%. Prioritize exploitation: small variations around the current best rather than broader exploration. We are particularly interested in identifying whether the optimum is robust to small changes in catalyst loading.
Summary
Expert context can be edited any time, including during a run.
Updates affect the next iteration the AI generates, not the current one.
Use the background portion to ground the AI in the problem; use a short "Current focus" section to steer the search as the run progresses.
Revise intuitions once data has tested them, so the AI doesn't keep optimizing toward a hunch that's been disproven.
The AI will not follow context that contradicts the data — your edits guide it, but the observations remain the final word.