4d • General discussion
Long Term Project
Working through a long-term project with my agent (Hermes on a Mac Mini, talks through Slack right now). Right now it's just me talking to her, but as I bring in more people on the ops side, I want a real way to control who can tell her to do what, not just whoever's in the Slack channel.
First piece I'm building: a permissions management UI sitting outside Slack. Before I start from scratch, curious if anyone in here has already solved this for their own setup. What did you build it on, and what would you do differently if you started over?
For anyone who wants context on what I'm going to build (slowly)
----
What this is
An internal tool for my own business, not a product I'm trying to sell. Alina is an AI agent running on a Mac Mini via the Hermes gateway framework, currently accessible only through Slack. This is a long-term, incremental build — roughly an hour or two a day, ongoing, no fixed deadline. Think of it the way you'd think about developing an employee: capabilities, access, and trust grow over time as Alina takes on more tasks, more integrations, and more data.
The actual problem
Talking to an agent purely through Slack conversation doesn't scale as a way to manage or trust it. Every instruction, every permission, every piece of context lives in scattered messages. There's no way to verify she actually did what was asked — only her own report that she did it. As she takes on more tasks, more connections, and more integrations, that gap gets worse, not better, because there's more surface area and no more visibility than before.
This is a verification problem first, and a control/visibility problem second. A nice dashboard showing what she's capable of doesn't fix it if it can't also show what she actually did.
The vision: a web control plane alongside Slack
Not a replacement for talking to her in Slack — a second surface that gives direct visibility and control, instead of relying entirely on her self-reporting.
It covers five things:
Skills and jobs. What she's capable of doing, and what she's actively running or scheduled to run. Read-only visibility into her current capability set.
Verification / audit trail. The most important piece. A record of what she actually did — tool calls, API writes, messages sent, outcomes — captured automatically by Hermes at the point of action, not narrated by Alina after the fact. The trust problem only gets solved if the log is written by the infrastructure, not by the agent describing itself.
Permissions and security. Who can instruct her to do what. Matters once more than one person is interacting with her directly — not needed preemptively while it's a single-user setup.
Integrations and data warehousing. The growing set of connections into whatever business systems I run on, and a single unified data layer underneath them rather than several disconnected sources.
Backup and versioning. Ability to roll back. Her skills/config are files — versioned the normal way, revert to any prior state. Her accumulated memory/state is different — that needs point-in-time snapshots, not git, since it's a growing log rather than an editable file.
Build philosophy
Incremental, prioritized by actual pain rather than a fixed roadmap. Verification (the audit log) comes before the cosmetic skills/permissions UI, since that's the problem actually being felt right now. Multi-user permissions get built when a second person is actually in the loop, not before. Data warehousing and integrations grow as connections are actually added, not built out speculatively in advance.
As the underlying models get better and Alina takes on more, the job of this platform is to keep visibility and control proportional to that growth — so trust in what she's doing doesn't have to rely on memory or her own narration of events.
0
1 comment
Long Term Project
skool.com/productivebot
Official home for ProductiveBot users! Get updates, share experiences, and connect with others who are optimizing their life with AI Assistants.
Leaderboard (30-day)
1
+11
2
+7
3
+2
4
+2
Powered by