Agent fleet

The fleet is my 24/7 agentic system that encompasses the entire ecosystem that I have built around coding with agents like Claude Code. My day consists of two main actions, describing new features or ruling on decision batches.

Describing new features is pretty self explanatory. I open a new agent session and we discuss some new feature or service that I want to add to my homelab. We have a few back and forths, I add any key notes about architecture for this particular feature or service and then the agent banks a series of work items to complete the work.

Decision batches are the results of agents doing work items and reaching decision "forks" where they are supposed to stop and bank a decision item with the relevant background. As part of this there is a LLM reviewer that gates decisions before they are stored. An agent must include all the required sections in a clear voice to be able to add to the decision queue. These are all stored in a PostgreSQL database that allows me to block decisions and have chains of work items that need to be completed in sequence.

I open a new agent session, ask for the agent to give me the decisions for a particular project it goes out grabs the decision from the fleet database and gives them to me in groups with a few initial options. I rule on those and that becomes my decision database which is being used to train the oracle.

This system has expanded to an extensive CI/CD deployment with a separate repo to prevent agents from disabling the tests while they are building new features. It has an MCP (Model Context Protocol) server that the agents use to bank items and search for relevant decisions to make sure they aren't asking the same question multiple times. It relies on my architecture-checks to ensure that the agents are making code that meets my quality standards.

The sheer number of agents that can be running at one time also necessitates the use of shared dev-hooks that ensure that all code is run and tested on warm lanes in the homelab instead of crashing my local machine.