Back to all work
— Project 02
Agent infrastructure

AgentOS

Enterprise Multi-Agent Platform

AgentOS is a control plane for running AI agents inside a company: any employee can spin up their own isolated agents — or a whole team — and govern roles, approvals, channels, memory, and runtime from one dashboard, without touching the runtime underneath. I designed and built the product, backend, operator console, agent runtime orchestration, and AWS infrastructure solo.

Role
Solo Founder + Lead Engineer
Period
2025 to present
Status
Production
FastAPINext.jsAWS CDKOpenClawQdrantmem0FargateAzure BotSecrets Manager
— Chapter 01
System shape

How the system fits together.

Click a block to zoom in
Agents run under a control plane with roles, approvals, memory, and isolated runtimes. Click any block to see how a piece works.
Fig. 01 — AgentOS architecture
— Chapter 02
Decisions and outcomes

The calls that shaped it.

  1. 01

    A control plane, not a wrapper

    It orchestrates the OpenClaw runtime itself — spinning up, syncing, and tearing down each agent’s isolated container — so people create and watch agents from the dashboard while the runtime stays invisible. That one boundary shaped everything else.

  2. 02

    Self-serve agents

    Any employee stands up their own agent and the right tools and skills attach automatically from an internal marketplace. Minutes, not a ticket to an engineer.

  3. 03

    Describe a team, get a team

    A “Platform” builder agent turns a plain-English request into a fleet — picking the agents, writing each one’s SOUL.md (role, voice, guardrails), and assigning skills, schedules, and channels — as a blueprint you approve first.

  4. 04

    Every write is gated

    A GateKeeper agent approves each external action before it lands, escalating to a human when it matters — and it fails safe: if the gate is down, the action is blocked, not waved through.

  5. 05

    Isolated by default

    Each agent is sandboxed with no path out except channels you approve. Memory is private-to-agent or shared-across-team (Qdrant + mem0); when agents need each other, they talk through explicit cross-agent plugins — collaboration as a feature, not a leak.

  6. 06

    Capability in one click

    Assign any skill or tool to any agent from a single marketplace — everyday ones self-served, sensitive ones admin-approved, each tracking who built it.

  7. 07

    An operations product, not a demo

    Self-service Teams bots with per-user allow-lists, live fleet monitoring, infra-vs-model cost tracking, full audit history, and a monitoring agent watching the others — all infrastructure-as-code.

— Aside
The interesting work isn't the stack. It's the boundaries.
— Chapter 03
How it runs

What it runs on.

Backend
Python · FastAPI
Frontend
Next.js · React
AI / Agents
OpenClaw runtimeQdrantmem0
Channels
Azure BotSlack · Telegram
Cloud
AWS FargateAWS CDKDockerSecrets ManagerGitHub OIDC