PR System for Persistent AI Agents
Filip Kozera, CEO of Wordware, on the rise of vibe doing
The hard part in AI knowledge work is not generating an answer, it is deciding what becomes durable truth. In software, git gives agents a clean review loop, a diff, an approval step, and a main branch. In a product like Sauna, the working memory is built from live email, chat, meetings, and docs, so the system needs a way to separate passing chatter from facts and preferences that should keep shaping future actions.
-
This is why persistent agents are architecturally different from one shot agents like Manus. Manus could snapshot and shut down a workspace, then restart it later. Sauna is built for back and forth interaction in under a second, which means the sandbox itself becomes the product and has to stay coherent over time.
-
The closest analog is less an IDE and more an operating memory for a chief of staff. Sauna is meant to absorb signals from inboxes, calendars, Slack, and docs, then proactively complete batches of work. That creates much more value than an email drafter, but it also makes permissioning and memory hygiene central to the product.
-
Competitors split on this design choice. Specialist tools like Fyxer narrow the problem to email triage and reply drafting, which reduces the chance that noisy context contaminates the system. Platform players like Claude Cowork also gate access through explicit folders and approvals, showing that review layers are becoming core product infrastructure, not just safety features.
The next wave of agent products will be won by companies that build a practical PR system for knowledge work. That means deciding what an agent can remember, what it can act on, and what needs human approval before it becomes part of the user’s ongoing operating context. The strongest products will feel less like chatbots and more like well run shared workspaces.