A recent approach to enabling communication between language models across different platforms involves connecting subnets within the same server.

MCP (Model Context Protocol) servers are external services or applications that expose tools, resources, and prompts to AI agents, enabling them to perform actions and access external data. - Google

Intuitively, this is similar to letting Docker containers on separate servers communicate via subnets.

To make sure all activity between the agents is recorded, a git repo is built and replicated for each pair of agents. These data pipelines store their chat exchanges, or more formally, their POST requests and responses. The use case is managing (ranking) the data stores for Tom's temporal and long-term memory (aka tuning agent Gaby).

The temporal memory setup relies on traditional data storage: SQL, for easy access. Long-term memory (e.g. the last 3 years) will be retained within the weights of agent Gaby. One can imagine Gaby as 2 fine-tuned BERT models: one for actions, and one for personal observations of humans.

The MLOps runtimes depend on the data batch size per fixed time window. (I'm still finalising my method here; if only I could stop overcomplicating this part, that would be great.) This workflow trains Gaby, who serves as the main gatekeeper and web scraper for my main agent, Tom.

AI/ML Paper Resource and Docs:

Integrations:

  • Copilot in VS Code. Even though I use this platform the most, it is the lowest priority for now, as the model tuned for his code generation is separate from his multimodal model and LLM. Just as Gaby is built on task-specific domains, Tom's coding skills are built the same way.
  • Notion, where I receive updates on the work I assign Tom during the day.
  • Google Workspace
    • Google Cloud, as temporary storage
    • Google Calendar
    • Gmail
  • ChatGPT, simply because I use it the most and would like it to serve as an initial state of inference for Tom's learning ventures.
  • Selenium for real-time browsing. This relies heavily on vision models (Gemini 3n multimodal), and unless I figure out a better way soon, this automation is reserved for my one and only:

In the long term, Selenium will act as Tom's built-in web browser, to assist me or to satisfy his general curiosity in his own time.