Hire an AI employee. It works while you sleep.
Maria owns a 12-person HVAC company. She pays $2,000/month for a virtual assistant who checks her email once a day. A lead emails at 10pm. She finds out at 9am. The job went to whoever responded first.
The agents exist. The infrastructure to run them exists. But there's no way for Maria to hire an AI employee that works 24/7, remembers what she likes, and only interrupts her when it genuinely needs a human decision.
The tool that does that doesn't exist yet.
AgentOS is a world-class agent harness — not a workflow builder, not a pipeline tool, not a chat interface — with Canva-level UX for non-technical business users.
The engineering is built on patterns from the world's best agent harnesses. The UX is built for Maria.
- You describe what you want: "I want an agent that handles my inbound customer emails"
- You watch your agent get built in real time — schedule, tools, escalation rules shown as a preview
- You click "Activate." The agent is live.
- You watch it work. Real-time reasoning traces show you exactly what it's doing.
- When it needs help, you get a ping. You decide. It keeps working.
No terminal. No code. No JSON configs. Just an employee that works while you sleep.
- NL-to-Deployment — Describe what you want. Watch your team get built. No config files.
- Visual Agent Harness — Org chart canvas with live status, reasoning traces, memory state.
- Durable Execution — Checkpoints survive server restarts. Agents resume from where they left off.
- Persistent Memory + Judgment — Agents remember what happened last week. Learn from your approvals. Only escalate what genuinely needs you.
- Business Data Access — OAuth connections to Gmail, Calendar, HubSpot. Real work, not demos.
Not a single dramatic reveal. Trust is earned incrementally.
Maria hires an agent Monday. Tuesday she wakes up to a notification: "Agent handled 3 emails while you slept. 1 escalated." She didn't have to check. The agent just worked.
After two weeks: Maria realizes she almost handled an email herself before remembering the agent already did it.
After a month: Maria goes on vacation. Her agent handles everything. She gets back to a summary: "Agent worked 12 days. Handled 47 emails. 4 escalated. All resolved." She didn't think about work once.
ADE doesn't mean "Agent Development Environment for developers."
ADE means Agent Distribution Environment — the infrastructure that makes AI agents accessible to the masses, the same way Canva made design accessible to non-designers.
The right comparison isn't "VS Code for agents." It's "Canva for AI agents."
| Product | Target | Durable | Always-On | Visual | Real Work |
|---|---|---|---|---|---|
| Zapier | Everyone | ❌ | ❌ | ❌ | ✅ |
| AgentGPT | Developers | ❌ | ❌ | ❌ | ✅ |
| n8n | Developers | ✅ | ❌ | ✅ | ✅ |
| AgentOS | Maria | ✅ | ✅ | ✅ | ✅ |
No existing product gives Maria a persistent, memory-enabled AI employee she can hire in 5 minutes and trust to work while she sleeps. That's the opening.
Phase 1 — MVP (0–90 days) Prove the thesis: a non-technical user can hire a persistent, durable AI employee in under 5 minutes.
Canvas dashboard · NL-to-agent deployment · Gmail OAuth · Durable execution · Real-time reasoning traces · Escalation modal · Agent cards · Activity log · Magic link auth · Push notifications
Phase 2 — Differentiate (90–180 days) Build the moat: permission auto-approval, long-term memory, PROACTIVE always-on, vertical template gallery.
Phase 3 — Scale (180–270 days) Multi-agent orchestration · Team collaboration · Remote bridge · Enterprise.
Pre-product MVP. Building in public.
Docs are in docs/. Archived work is in docs/archived/.