Inspiration
The most common knowledge-management problem inside companies is this: experience lives in people's heads, data is scattered across disconnected systems, and process knowledge only gets passed along by word of mouth. Every time an employee runs into a problem, they have to ask around, dig through documents, or guess at the "right" way to do something — while managers have zero visibility into any of it, and nothing gets captured or governed. We wanted to bring three things — enterprise knowledge, operational data, and business actions — into a single Agent entry point: employees get a natural-language interface to ask questions and take action, managers get a governable control plane, and the company's knowledge and data stay inside its own system boundary instead of being scattered across individual chat histories. ourproduct is that idea, built.
What it does
an enterprise-grade Agent workspace: Employees can ask questions, look up data, and trigger business actions in natural language, without switching back and forth between multiple systems On the management side, there's visibility into what the Agent was asked, what it did, and what data it accessed — creating a traceable, governable record All data and workflows stay inside the company's own systems and never leave
How we built it
Chat Completions API
Challenges we ran into
Reliable tool calling, multi-turn context management, and preventing unauthorized data access
Accomplishments that we're proud of
Two things surprised us in a good way. First, response speed — even with a fully custom agent orchestration layer on top of the raw Chat Completions API (no Assistants/Agent SDK), the end-to-end latency stayed snappy enough that the interaction felt closer to a normal chat than a slow enterprise tool. Second, multi-turn stability — we'd expected context to degrade or drift over longer conversations, especially with tool calls mixed in, but it held up better than we anticipated, keeping track of what had already been asked and what data had already been accessed across multiple turns.
What we learned
Going in, we assumed "governable" mostly meant access control — deciding upfront who could reach what data. What we actually learned building VectA is that the harder, more important problem is traceability: giving managers a clear after-the-fact view of exactly what the Agent was asked, what it called, and what data it touched. Access control alone doesn't make a system governable if no one can see what happened — visibility has to come first, because you can't govern what you can't see. That shifted how we think about the whole product: the logging and audit layer isn't a secondary feature bolted onto the Agent, it's actually as central to "governance" as the permission checks themselves.
What's next for One Agent Workspace for Enterprise Knowledge & Data
The next step is scaling validation — taking it into a real enterprise pilot to see whether it holds up under real data volume and real-world scenarios, not just the curated flows we tested during the hackathon. That means testing with messier, larger datasets, more concurrent users, and the kind of edge cases and ambiguous queries that only show up once real employees — not just us — are using it day to day. If the traceability and multi-turn stability we saw in the demo hold up under that kind of load, that's the real proof this approach can work at enterprise scale.
Log in or sign up for Devpost to join the conversation.