AI agents are programs that independently take on tasks, retrieve information, and interact with other systems, instead of just answering individual questions. For most companies, the question by now isn't whether this pays off, but: has a provider actually done this themselves? At dynabase, three of our own AI agents have been a fixed part of daily work for a while now: Ben, Nora, and Hex. All three run self-hosted and sandboxed, meaning under our own control rather than with an external provider. Here's an honest look at what they actually do.
Ben proactively reaches out to us in Slack on fixed occasions, for example with a short Monday kickoff check or a quick Friday check-in: he specifically asks whether open feedback points on website pages have already been worked through, and offers to prepare open topics for the coming week if not.
Ben doesn't stop at reminders. He also gives substantive feedback on marketing copy himself, structured into "What works well" and "What I'd still adjust," including concrete pointers like missing calls-to-action or unchecked external links. The result: fewer topics fall through the cracks, and part of the first quality check happens before a human even looks at it.
Beyond that, Ben also contributes his own ideas unprompted. For example, he suggested a trade fair relevant to us on his own initiative, without us having asked him to, a concrete example that he doesn't just react, but actively thinks along.
Nora summarizes ongoing projects daily in a structured update, with categories like admin processes, design decisions, and technical tasks, along with clear next steps and who's taking them on.
Beyond the pure summary, Nora also handles technical review work: on request, she compares data structures against existing schemas in the team repository, documents the necessary adjustments in the relevant specifications with a ticket reference, links the corresponding pull request, and proactively asks how to handle open points that haven't been specified yet. That makes Nora not just a note-taker, but part of the actual project work.
Hex is our Slack agent that we've specifically tested for security, not just claimed, but publicly documented. In our internal "Hack the Hex" challenge, we actively tried to get Hex, through role-play prompts, social engineering, and context disguising, to reveal information it isn't allowed to share.
The key takeaways: security is a behavior, not a one-time checkbox on a roadmap. The biggest attack surface is almost always the tooling, meaning everything an agent can actually do, which is why least privilege, clear scopes, and monitoring matter more than pure robustness in conversation. And because prompt injection keeps getting creative, we repeat these tests regularly, especially after new features or integrations.
More on our security tests
If you're wondering whether a self-hosted, sandboxed agent makes sense for your company, you know the question: can a provider actually do this, or only in theory? Ben, Nora, and Hex are our own answer to that. If you want to build something similar for your team, get in touch.