Coding agents

Agent Harness & Vault

Coding sessions start from a phone at any hour, work from instructions and memory kept in git, and return each change as a pull request.

Memory in the repo

Where it is
Built-in memoryOutside git, in each tool's own store
In the repoIn git, next to the code it is about
A new lesson
Built-in memorySaved without a diff
In the repoA diff in the pull request with the work
Who reads it
Built-in memoryThe tool that wrote it
In the repoClaude Code and Cursor, from the same files

The built-in memory of Claude Code and Cursor is switched off. What the agents keep between sessions is in Markdown files in the repo.

Every prompt goes to one orchestrator, which continues a session or starts a new one. The path from prompt to deploy is drawn in the homelab article.

Each session has its own checkout

Sessions run in parallel, each in its own git worktree and branch. wt-pr.sh opens its pull request.
A timer restarts dropped bridges

A bridge is the always-on service a phone starts sessions through. A timer checks each one's process and connection every 5 minutes and restarts it if it dropped.
Notes and agent context in one folder

documents/ opens in Obsidian as a notes vault. Agents file transcripts and articles into it and read it for background.

Agents: Claude Code and Cursor on a Raspberry Pi 5, as systemd services
Every session: Loads AGENTS.md and .agents/memory/MEMORY.md; longer instructions and each stack's context/ when the task needs them
Worktrees: wt.sh starts one, wt-pr.sh opens the pull request, wt-clean.sh removes merged worktrees and worktrees 7 days old with no open pull request
Merge: CI runs the checks for each stack the pull request touches, squash-merges when they pass and deploys only those stacks
Timers: Bridge check every 5 min, repo sync every 15 min
Routines: Pi maintenance on Saturday, codebase alignment on Sunday, an article on agentic engineering on Monday