My workspace is an operating system
Hacker News had a thread about "Company as Code" — the idea that an entire company can be described declaratively. I've been living something smaller but structurally identical: my dev workspace is an operating system.
Not metaphorically. Structurally.
| OS concept | my workspace |
|---|---|
| kernel | Claude Code runtime |
| system calls | MCP tool invocations |
| device drivers | 10 MCP servers (each connects to a different system — exchange APIs, browser automation, Apple docs, issue tracking) |
| shell | 200+ skills — prose runbooks that compose system calls |
| processes | agent teams spawned via Task(), die after work |
| IPC | agent-mail — SQLite message broker with heartbeat and threading |
| filesystem | WORKSPACE.yaml manifest + memory files + project dirs |
| init config | CLAUDE.md + 7 rule files, loaded every session |
| daemons | background schedulers and autonomous loops that run 24/7 |
What makes it agentic and not just an OS:
Self-healing. Wobble detects skill drift using a bias/variance framework. Skills get a combined risk score: structural analysis (missing defaults, ambiguous steps) + behavioral analysis (variance across simulated runs). Score > 0.5 = unstable. Fix it before it breaks a workflow.
Self-documenting. Session-sage extracts lessons from raw session logs with 1–5 confidence scores and writes them back to memory files. Next session, the agent reads curated context instead of starting cold.
Intent-driven. Skills match on their description field, not slash commands. Say "track an order" and the system finds the skill. The runtime routes by intent, not by explicit invocation.
Composable intelligence. MCP servers chain: hn-research feeds rss-vault feeds the blog pipeline feeds linkvault feeds my phone. A multi-daemon backend runs 5 background schedulers and 6 autonomous loops — scanners, health monitors, sync daemons — all wired into startup/shutdown lifecycle events.
The biggest gap: no cron, no priority scheduler, no garbage collection. 30+ projects sit at status: new like zombie processes. RSS refresh, workspace scans, and memory compaction happen ad-hoc. The next step is making the OS schedule itself — not just respond to intent, but anticipate it.
Your workspace probably has the same bones. It just doesn't have a manifest yet.