Made a multi-agent AI show its work, so scientists could trust and act on it
CHALLENGE
Scientists could not tell how BenchSci's co-pilot, its conversational AI assistant, was working, what it had searched, or why an answer looked the way it did. This resulted in them hesitating to act on what it returned.
WHAT I DID
I worked with engineering as they rebuilt the co-scientist as a multi-agent system, and designed the pattern that made its activity legible and traceable without overwhelming scientists.
OUTCOME
The same scientists who'd called the product broken reported the opposite after the pattern shipped: more visibility into the agent's reasoning, and more trust in, and control over, the decisions they made from it. The pattern outlived the feature. It's now the default for how BenchSci's AI explains its reasoning across its co-pilot product.
ROLE
Principal Product Designer. Sole designer for the co-scientist experience.
TIMELINE
Jun – Aug 2026
TEAM
PM, five engineers, project manager, three scientists.
DETAILED PROJECT BREAKDOWN
Problem: an invisible process broke trust and control
In 2023 I set trust through visibility as BenchSci's operating principle for AI: scientists trust an answer they can verify quickly and locally.
Read about it here →
By 2025 BenchSci's flagship product was a conversational AI co-scientist searching scientific literature, clinical trials, and proprietary data to answer a single question. It also got a lot harder to watch work.
First instinct: show everything. Wrong instinct
The first version streamed a raw feed of every agent event as it fired. Essentially a log. Multiple searches running at once, each reporting its own status, in real time.
The logic was straightforward. Scientists are skeptical, scientists want evidence, so show them everything.
It failed for three reasons.
It was overwhelming. Multiple concurrent searches produced more events than anyone could read at the speed they arrived. The feed moved faster than comprehension.
Raw system updates were not written for humans. They described what the system was doing in the system’s terms. Legible to an engineer, noise to a scientist.
Progress was non-linear. Agents do not proceed in a straight line. Exposed raw, that looked like the system doubling back or thrashing, which made it appear less competent than it was. Showing everything made a working system look worse than showing nothing.
Feature walkthrough: what shipped
One constantly updating reasoning section that is expandable into the full trace with a click.
Decision: default to one line, expand to the full trace
Stream updates live, with progressive disclosure.
Default state shows only the latest meaningful update: one line, human-readable, always moving.
Expandable to the full process, one click away, for anyone who needs to inspect what was searched.
That split solved both problems with one pattern. The default answers "is it working." The expanded view answers "what did it do," which is the part scientists and enterprise evaluators needed for the record.
The trade-off: the expanded view still shows relatively raw activity. I knew going in that a better version existed, and shipped without it because of the two-month timeline. The default state was doing the heavy lifting, so raw detail behind a deliberate click was an acceptable cost. That was the right call to ship.
Outcome: scientists went from guessing to describing what it was doing
Before. 30 to 90 seconds of near-silence on the flagship product’s core interaction. Scientists waited, some left, and the consistent report was that it felt broken. No way to see what had been searched, so no way to document the method or answer an evaluator’s question about it.
After. A continuously moving, human-readable account of what the system was doing, with the full process one click away.
In moderated sessions, scientists could describe what the system was doing while it worked. In the previous version they could not tell running from stuck.
QA-athon with scientists reported less confusion during long-running responses.
Customer feedback was positive.
The process view answered the inspectability question enterprise evaluators had been asking.
Where I wanted to take it
Group the low-level agent events into higher-level progress beats, so the expanded view reads as clearly as the default line does.
Preserve full visibility while cutting the noise further, so inspecting the process never means wading through a raw log again.
That's the version I'd have shipped with more time. The two-month window meant choosing between a good default and a good detail view. I didn't get both.
What I'd instrument next
Abandonment during long-running responses. The direct measure of the original problem. Does the user believe it is broken, or does the user believe it is thinking?
Expansion rate on the detail view. If almost nobody expands, the default is too shallow. If almost everybody does, the default is hiding something they need. I set that default from qualitative sessions and never got to validate it.
Time to first action after the answer lands. Hesitation is the tell for uncertainty.
The principle outlived the project
2023, evidence traceability: can I check where this claim came from?
2025, process legibility: can I understand what the system did to get here?
Ongoing: this pattern is now the default for how BenchSci's co-scientist communicates agent reasoning, and it carried into EMET, the company's current expert-agent platform.
A principle set in a week of discovery research in 2023 was still doing structural work three years and one architecture change later, because it was defined at the right level of abstraction and documented as a primitive, not a screen.
