RoyOS

RoyOS Devlog #4 — Making the Runtime Safer

This round of development was focused on making RoyOS more reliable as I expand what it can actually do.

I strengthened the tool execution and approval system so that all 70 registered tools have explicit effect classifications, and unknown tools are blocked by default. I also added adversarial tests for approval bypasses, replay attempts, modified inputs, and shell-based workarounds.

I also integrated five additional product workflows into the existing project ecosystem, including tools for content creation, messaging, and marketing. These remain in mock mode while I continue validating their behavior and security. On the dashboard side, I hardened API routes, file access, uploads, security headers, and rate limiting.

Another area of progress has been the website and mobile app builders. I’ve added a reusable motion-design system to the website factory and validated the mobile app builder locally.

The biggest takeaway is that adding capabilities is only part of building an autonomous system. Each new tool creates more opportunities for failures, so permissions, approvals, state management, and regression testing have to improve alongside functionality.

Next, I want to close the remaining security findings and improve how RoyOS coordinates complete, multi-step workflows across its tools.

RoyOS Devlog #3

RoyOS is starting to feel less like a prototype and more like an actual operating system for autonomous work.

This week I focused heavily on reliability and the boundary between Jarvis and the RoyOS runtime.

Jarvis now handles conversational tasks directly through Claude, while RoyOS acts as the durable background worker for longer-running jobs. That separation has made the whole system much more predictable.

I also tightened up the RoyOS bridge and safety system. Jobs are forced into a scratch environment during testing, live writes require explicit authorization, and destructive operations still require human approval.

The runtime is now passing all tests, and the bridge has been heavily regression-tested as well.

The biggest lesson from this stage has been that an autonomous system doesn’t just need intelligence. It needs state, provenance, permissions, retries, observability, and clear failure behavior.

There’s still work to do, especially around getting more complex multi-step objectives to complete reliably, but the foundation is becoming much stronger.

Next up: continuing to push RoyOS toward a system that can take a real objective, break it into durable work, execute it safely, and report the result without needing constant supervision.

Devlog #2

Recently, I have been focused on turning RoyOS into a more capable autonomous agent runtime.

I built out multi-agent delegation, allowing agents to create and delegate tasks to other specialized agents while keeping permissions and delegation depth strictly bounded.

I also improved agent creation and validation, so dynamically generated agents have to pass validation before they can run.

A major focus was reliability and safety. RoyOS now supports persistent runtime state, task recovery after crashes, structured execution events, replanning, and approval states.

I also found and fixed a filesystem security issue involving dangling symlinks that could allow an agent to write outside the allowed directory. I added regression and mutation tests to make sure the fix actually holds.

The test suite has now grown to passing many tests with TypeScript compiling cleanly.

Next, I’m working on getting RoyOS to reliably discover and use the actual agent definitions before moving toward production use.

Set up the core repository and development workflow for RoyOS, an agentic operating system combining a dashboard interface with an automation runtime. Structured the Next.js dashboard project, cleaned up the codebase, and prepared it for deployment.
Connected the project to GitHub and configured version control for ongoing development. Deployed the dashboard to Vercel to establish a live production environment and enable instant previews for future updates.
This session focused on foundational setup, deployment pipeline, and preparing the system architecture for continued development of agent-based workflows and runtime integration.