Ultron — Development skill for Claude Code

Ultron

**An agent that creates the harness that creates itself in various environments and adapts.**

Ultron is a modular, self-evolving **harness ecology** built around a *preserved* [`hermes-agent`](https://github.com/NousResearch/hermes-agent) core. You just chat. The agent builds the tools you need, renders them as **generative UI inside the chat**, and the more you use it the more your harness is **whittled to your workflow** — automatically, safely, and reversibly.

---

Philosophy

• Preserve the core. Upstream Hermes stays untouched behind explicit adapter seams. Ultron never mutates Hermes source, memory, skills, tools, or config — it only attaches to it.
• Modules are harnesses. A centralized agent creates modules (prompt packs, tool policies, UI panels, safety/budget) like tool-calls. Those modules are your harness.
• Use shapes the harness. Your usage, request history, and feedback become the selection pressure. Modules that help survive; modules that don't atrophy. Like a stone shaped by a river.
• Generative UI is the surface. Tools and results appear as typed, server-validated UI components rendered inline in the chat — you see your work happen, you don't operate a dashboard.
• Safety is non-negotiable. Self-modification is bounded (one change at a time), evidence-gated, reversible (no-poisoning rollback), audited, and privacy-redacted by construction.

What you experience

• A command surface (/). A command bar drives a bounded generative-UI canvas (not a scrolling transcript): ask for what you want and the agent builds/tunes a tool, streaming the result back as generative-UI cards (plan, risk, tests, tool result, evidence, safety). Toggle Replace (A) vs Accumulate (B) to transform the canvas or keep a capped workspace; pin cards you want to keep. Attach an image to drive the multimodal (VLM) path. 👍/👎 feedback quietly shapes your harness over time.
• **A settings dashboard (/dashboard)