Anthropic has a philosopher named Amanda Askell whose job is, in part, to think about Claude’s personality. She treats it as a character question: What does it mean to be good, and how should Claude act accordingly?
Most teams building agents don’t have an Amanda Askell. Agent personality is often an afterthought or the byproduct of whatever the underlying model or system prompt produces. When we started designing Global Agent, we wanted to do better than that, so we decided to treat personality the way we’d treat any other product surface. Pick the behaviors that matter, define what good looks like, and fine-tune when they’re off.
There are two traits that stood out to us as having the biggest impact on whether someone would keep using Global Agent:
These are certainly not the only personality dimensions worth tuning, but they’re the ones we started with.
When we first launched Global Agent, we gave it a heavy bias for action. The system prompt instructed the agent not to ask follow-up questions in response to an initial user prompt.
The reasoning made sense at the time. We assumed users were coming to Global Agent without knowing their data taxonomy and event properties, so asking them clarifying questions upfront would make them bail. We thought it’d be better to let the agent take a swing and then have the user course-correct.
That turned out to be the wrong instinct. The agent was overconfident in places it shouldn’t have been, made incorrect assumptions, and didn’t accurately answer user questions. It needed a personality change.
We decided to A/B test an agent that paused to ask clarifying questions against the original. Users who interacted with the inquisitive agent were more likely to have longer conversations and save its analysis. That was enough for us to dial inquisitiveness up.
While we started with a reasonable intuition based on user behavior, it was ultimately just a hypothesis. And whenever you have a hypothesis about agent personality, you need a way to check it against what your agent is actually doing in real conversations.
We noticed an interesting pattern showing up in our internal dogfooding channels. Global Agent would tell users how to do what they were requesting, rather than acting like a helpful colleague and offering to do it for them.
Our instinct was to tune up the agent’s eagerness to help, but we weren’t sure how prevalent this issue really was. A few people in Slack might not translate to hundreds of users, so how do we know if this is a real issue?
This is exactly the kind of question that we built Agent Analytics to help answer. Anthropic invests heavily in interpretability research, which involves opening up the model to understand how it thinks. Agent Analytics comes from a similar instinct, but one layer up, to understand agent behavior and performance.
We created an Agent Analytics evaluator that reads every Global Agent conversation and flags any time the agent gave instructions when it could have offered to do the task itself. The percentage was significant enough (~3% of all conversations) that we dialed up the agent’s helpfulness trait.
User feedback gave us a helpful starting point, but sizing the issue with analytics is what gave us the confidence to make the change.
Start with a hypothesis about how your agent should behave and identify the traits needed to achieve it. Ship your agent’s personality and, like any new product feature, listen to what users say and measure the results.
If something is off, decide whether it’s worth changing based on the size of the issue, then A/B test the fix. The first part is where most teams get stuck. Without a way to measure how often a personality issue is showing up, you’ll end up fine-tuning based on the loudest feedback you hear.
For any team building agents, that’s the part worth investing in: a way to turn the constant stream of issues into something you can size and prioritize. We built and use Agent Analytics to do exactly that.
Inquisitiveness and helpfulness are only two pieces of the personality puzzle. Other traits, like warmth, formality, and directness, help make up the whole picture.
Even within inquisitiveness and helpfulness, there’s an entire spectrum where we could have landed, but no single default fits everyone. Some users loved that we dialed up Global Agent’s inquisitiveness, but others were frustrated by it pausing to ask questions. There’s no version that’s right for everyone; there’s just a default that has to pick a side, and a way for the other side to override it.
That’s what we’re working on next: a personality editor that lets users define the agent’s personality and fine-tune its traits to fit their preferences. We’ll share more when it’s shipped.
Most teams ship their agent personalities by accident. We wanted to design ours on purpose.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。