Loading…
← All results// Case study 02 · Agent OS 2.0 · hybrid · production
Agent OS — supervised agent workforce
Agent OS 2.0 · hybrid · production
My own multi-repo business stack — orders, content, CRM — coordinated by a VPS control plane that runs 24/7 while code execution stays on a supervised local runner.
Prod smoke PASS — HITL approve/reject/cancel, Langfuse cost tracking, E2E handoff on disk. Guided demo ready; not a public SaaS.
Before — Problem
- ✕ Orders, content and CRM scattered across spreadsheets
- ✕ Same questions answered in five different tools
- ✕ No single place for rules — everyone improvises
- ✕ Deploys feel risky because nothing is documented
After — Resolution
- ✓ Orchestration and HITL run 24/7 on VPS — queue, approve, reject, cancel
- ✓ Agents run in fixed roles with clear contracts (5-node LangGraph)
- ✓ Code and git work on a supervised local runner — not 24/7 without your PC
- ✓ Human approval before anything hits production — no auto-deploy
Architecture
FastAPI + LangGraph + PostgreSQL on an EU VPS. Hybrid routing via WAITING_RUNNER: orchestration and HITL online; git/code work on the dev PC. Mission Control UI for tasks, queue, history and costs.
Not built by one prompt. Assembled by agents that plan, code, test and review — with a human at the gate.
LangGraph pipeline
Planner
Reads SSoT, breaks work into tasks and context packets.
Coder
Implements one module per session — least privilege.
Tester
Runs tests and collects logs before review.
Reviewer
Human gate — LangGraph pauses at REVIEWING.
Summarizer
Closes the session with handoff markdown on disk.
VCMS governs repos and SSoT — outside the LangGraph pipeline, not a graph node.
The Build
Five delivery phases — same as How I work on the home page. Client-facing workflow, separate from the LangGraph pipeline above.
01
Map
Find leaks, score ROI, agree first system. Deliverable: Automation Map.
02
Architect
Target diagram + spec — who does what, where data flows.
03
Build
AI workforce assembles against fixed rules in test environment.
04
Verify
Real scenarios + your sign-off. Gates built into the system.
05
Hand over
README, walkthrough, optional care. No lock-in.
Pipeline map
Planner → Coder → Tester → Reviewer → Summarizer — each step has a contract before anything ships.
Planner → Coder → Tester → Reviewer → Summarizer — hybrid pipeline with human gate before ship.
What is LIVE today
Guided agency demo — not a public SaaS. Hybrid control plane on VPS; code execution stays on a supervised local runner.
- Task queue + lifecycle LIVE
- HITL approve / reject / cancel LIVE
- Hybrid VPS + local runner LIVE
- Code/git execution LOCAL-ONLY
- Langfuse cost tracking LIVE
- MCP gateway (filesystem/git) LIVE
- Demo 60s video LIVE
- Handoff pipeline E2E LIVE
- Task cancel UX LIVE
- Local runner OpenCode LOCAL-ONLY
- Public multi-tenant SaaS NOT OFFERED
- Auto-deploy to client PROD NOT OFFERED
LangGraph pipeline: 5 nodespytest integration: 42/42prod E2E smoke: PASSbrowser smoke: 2026-06-24
Mission Control
Tasks, queue, history and cost tabs — observability layer for Agent OS. Login required for live UI; screenshot from prod smoke.
Mission Control dashboard — tasks, queue, history and cost tabs
Mission Control — tasks, queue, history and cost tabs. LIVE on os.flexgrafik.nl.
Agent cards — sample contracts
Each execution agent has a fixed role. No mega-agent, no scope creep — least privilege by design.
| Agent | Scope | Output |
|---|---|---|
| Planner | SSoT read, task list | Implementation plan |
| Coder | Target module only | Diff ready for review |
| Tester | Tests + logs | Pass/fail report |
| Reviewer | Human gate (HITL) | Approve or block |
| Summarizer | Handoff + session closure | Handoff markdown on disk |
Stack
FastAPILangGraphPostgreSQLMission ControlEU VPS
Proof
Agent OS agent cards — five fixed roles
Every agent has a role, rules and a review gate — LIVE on VPS; code execution LOCAL on dev PC.
Measurement
5-node LangGraph · hybrid VPS control plane LIVE · prod E2E handoff · Langfuse cost tracking.
Could this work for your business?
Start with a paid Automation Map. In 60–90 minutes we map your exact setup, score the ROI, and recommend whether this is the right first system — before you commit to anything bigger.
Book your Automation Map