You find out it's wrong from a customer
Without a score or a test suite, the first person to catch a bad answer is whoever acted on it.
Agents
An analyst you can hold accountable
Each agent carries a confidence score, an evaluation suite it has to keep passing, and 7 tuning surfaces — identity, knowledge, memory, rules, tables, verified queries and skills — under your control.
Acme’s Workspace
Search⌘ + K
John Doe
Revenue Analyst
Created by John Doe · 1,284 questions answered
Trusted by 1,000+ businesses across the world
An AI you can't test is an AI you can't trust
Tuning is a prompt someone edits blind
One text box, no measure of whether the edit made the answers better or quietly broke three of them.
It never learns what you corrected
You fix the same misread column every week, because the fix lives in a chat thread instead of the agent.
Tune it, watch it work, score it
7 dials, not one prompt
An agent is more than its instructions. Who it is, what it knows, what it remembers, the rules it obeys, the context on each table, the queries you have signed off and the skills it can load are each their own surface, so a fix lands where the problem is.
- Identity, knowledge base and memory
- Rules and per-table configuration
- Verified queries and reusable skills
Watch it do the work
An agent reads the sources you gave it and writes back what it learned. Every step it takes is named as it happens, so the run is something you read rather than something you trust.
- Each step named while it runs
- The documents it opened, listed
- What it learned written into its knowledge base
Evaluations that keep it honest
A suite of your real questions runs against the agent, each one scored and marked passed or failed, and the results roll up into one accuracy score. A regression shows up as a failed eval, not as a wrong answer in front of your team.
- Your own prompts as the test suite
- Pass or fail, with a score per prompt
- One accuracy score across the suite
What a tuned agent looks like
2K+ Agents trained
By the businesses running Supaboard in production.
97.8% Accuracy a tuned agent holds
Measured on a customer agent in production, after tuning.
7 Surfaces you can tune
Identity, knowledge, memory, rules, tables, verified queries, skills.
Frequently asked questions
What is an agent in Supaboard?
An AI analyst you configure and own: it has an identity, scoped access to your sources, and its own knowledge, rules and memory. Every analyst your team asks is one.
What is the confidence score?
A per-agent score out of a hundred, computed from its evaluation results and open issues. It moves when you tune the agent, so you can see whether a change helped before anyone relies on it.
How do evaluations work?
You collect real questions into a suite and run it against the agent. Each prompt gets a score and a pass or fail, and reruns track the trend, so a regression is caught by a failed eval instead of a user.
What can I actually tune?
7 surfaces: who the agent is, its knowledge base, what it remembers, the rules it must follow, per-table context, the verified queries it reuses, and the skills it can load.
Does an agent learn on its own?
It keeps a memory of what it learned from past conversations, and with self-improve on it proposes its own fixes — but learnings stay reviewable, and verified queries only enter the pool when you sign them off.