I've been working on many projects at once, running many subscriptions (more than two x20 Claude) and often ended up copying my co-workers questions to LLM in VSCode.
So I decided to move my work into a docker, isolated, hostable anywhere - locally or on server - and let my agents live there, stay live-connect to various channels and work for me and respond to my co-workers questions 24/7.
The container of agent contains API exposing it's file system, agentic sessions, git stuff and comes up with system of self-defining dockerfiles, so agents, when they need to get C++, Rust, ffmpeg or some other toolchain, will simply give you new dockerfile layer for approval and ask you to run rebuild.
This way agent's environment grows in predictable way and is replicable (unlike the local one), easy to move around, nothing is hidden.
Since then I've baked in robust tools to improve LLM efficiency (that are opt-in/out) and their impact on token usage is measured in your codebase.
Agents living in their container IDE can even be connected to CI/CD to act in real browsers on user stories kept in their file-system, verify an artifact and return their results to CI/CD as a quality gate.
It's fully extensible and pretty much everything UI you can find on website is just a nice UI derived from agent's file system - "a window" for a human operator.
For many weeks I already use Intentic to build Intentic. And I'm probably biased, but the experience is from another world...
Desktop app is about to be available soon too, that's my main focus now, although browser experience is really good to me!
I'll be very grateful for any feedback.