Introducing Findings: AI Agent Insights 🤖

Hey, guys.

So I wanted to share a new feature we're launching called Findings.

It comes from a pretty familiar problem.

You know, your AI agent can improve, but you really do not have time to crawl through every conversation and figure out where it's falling short.

So Findings, does that first pass for you?

It uses AI to review and categorize completed conversations, creating structured signals, like a user expressing frustration, an unresolved request, or an incorrect answer.

It then groups those recurring signals into findings, which can then be investigated by the Kapso Agent to understand what's behind them.

Let me show you how it works in your Kapso Dashboard.

Go to the Findings item in the sidebar, and here you'll be greeted by our onboarding.

When you first turn on Findings, it sets up two components.

First, it sets up a default event value.

This reviews completed conversations and creates structured signals like user frustration, unresolved requests, and incorrect responses.

Then it sets up a Kapso Investigator agent, which you can then run when a finding is created to investigate that pattern for you.

When you set it, up, you can also choose to review up to 50 conversations from the last seven days to start getting your first findings.

After gathering some data, we have our first finding, a recurring user frustration.

Because this is a recurring pattern rather than a rising trend, there's no baseline comparison.

However, the important thing is that this issue has continued appearing across multiple recent conversations.

So now, instead of manually dragging through dozens of conversations, I can open this finding and see the evidence behind it.

From here, I can also click on Investigate with AI or copy as a markdown file and hand it over to one of my coding agents.

I'm going to investigate with AI.

In this case, this launches the Kapso agent, which will then look at my workflows and execution logs.

And if I have a code base connected to Kapso, it can also inspect the code base to help find the root cause behind why this pattern keeps appearing.

Once the investigation is finished, the Kapso agent returns with suggested fixes that are specific and tailored to what I already have set up.

So for example, in this case it's suggesting that I add return and lookup failure rules to the agent.

I can look at the evidence in here and I can also review all the supporting evidence over here as well, which includes events, workflow, execution logs and even access to my code base.

If I have set up my code base inside of Kapso, It also tells me what it wasn't able to verify.

So I can also take that into account before making a change.

Now I can take these findings and these suggested fixes and apply them to my agent or workflow to fix them.

Once I'm done, I can mark the finding as addressed.

From there, findings will monitor new activity and look for signs that the problem is returning or whether the fix was actually useful.

Findings help surface the problem.

The evaluator creates and groups signals and the Kapso agent tells me what to change.

So going from something feeling off to specific evidence backed improvement is much easier now and I don't need to read every conversation myself.

Hope you guys like the feature and if you have any feedback please let us know because we're constantly working on improving it and tweaking it as we go along.

Thank you so much.