Senior AI Agent Engineer — Build & Run Our Autonomous OpenClaw Stack

Senior AI Agent Engineer — Build & Run Our Autonomous OpenClaw Stack
Hours: Full-time (40 hrs/week) preferred, will consider part-time for the right person
Salary: $800–$2,500/month based on experience and demonstrated capability
Type: Long-term, ongoing role. Starts with one defined build project. If you crush it, this becomes an ongoing position with substantial work for 12+ months.

What we're building
We run multiple businesses (commercial AI video production, real estate AI video platform, medical consulting clinics across multiple states) and we're consolidating our AI infrastructure into a single, powerful OpenClaw-based agent stack that we can drive from Telegram and use to autonomously ship real software, manage operations, and execute long-running projects.
We have most of the foundation already running on our Elestio VPS — n8n, Qdrant, Redis, Open WebUI, 13 MCP servers, GitHub org, Vercel integration, and an existing ---------- Telegram bot. We have active coding plans on Kimi, MiniMax, GLM (z.ai), and Qwen — all four of which have idle API capacity we want to fully utilize.
We are looking for one expert to architect and build the next-generation version of this stack with us. You will not be guessing — we know exactly what we want built and have a detailed architectural plan. We need someone who can execute it at a senior level, advise on tradeoffs, and ship working software at the end.

The starting project
A fresh, production-grade OpenClaw deployment that:

Routes all model calls through a LiteLLM gateway with task-aware routing across Kimi K2.6, GLM-4.6/4.7, Qwen3-Coder, and MiniMax M2 (planning ? GLM, coding ? Qwen, long-context ? Kimi, cheap/fast ? MiniMax)
Integrates cleanly with our existing Qdrant, n8n, Postgres, Redis, and current MCP servers
Connects Gmail, Google Drive, browser automation (Stagehand v3 + Browserbase), Bright Data, ---------- , and Composio connect-apps
Has the agent harness layer installed and configured: obra/superpowers, alirezarezvani/claude-skills (C-level + business-growth + marketing bundles), Anthropic's official skills, frontend-design
Has multi-agent observability working: disler/claude-code-hooks-multi-agent-observability and Langfuse self-hosted
Is architected with lethal trifecta awareness — clean separation between agents that read external content and agents with outbound write access
Supports parallel agent execution patterns (git worktrees, Agent Teams, Ralph loops where appropriate)
Is driven from Telegram (and ready to extend to WhatsApp via Meta's Business Cloud API)
Proves it works by being able to autonomously ship a real piece of software end-to-end — we'll pick the project together once the stack is up

We're not asking you to invent any of this. The architecture is documented. We need someone to execute it cleanly, advise where tradeoffs come up, and own the build.

What you must have (hard requirements)

Hands-on production experience with Claude Code, Cursor, Cline, Codex, or OpenClaw. Not "I read about it" — you've shipped real work using these tools.
Demonstrated experience with MCP (Model Context Protocol) — building servers, configuring clients, troubleshooting auth flows.
Experience routing through LiteLLM, OpenRouter, or Portkey with multiple OpenAI-compatible providers (Moonshot/Kimi, DashScope/Qwen, z.ai/GLM, MiniMax).
Strong Docker, Linux server admin, and Elestio/VPS management skills.
Experience with n8n, Qdrant, Postgres, Redis in production.
Familiarity with the harness engineering / context engineering discipline. You should know what SKILL.md, CLAUDE.md, progressive disclosure, subagents, and Agent Teams refer to without me explaining them.
Real understanding of AI agent security — prompt injection, the lethal trifecta, sandboxing, secrets management.
Strong English communication and async work discipline. You will document everything.

What's a strong plus

You've installed and used obra/superpowers yourself
You have published skills, MCP servers, or plugins on GitHub
You've built with Browser Use, Stagehand, or Browserbase in production
You' ---------- run agents through Telegram bots before (grammY, python-telegram-bot, or OpenClaw's gateway)
You understand the difference between Letta, Zep, and mem0 and when to use each
You've watched IndyDevDan, Cole Medin, Charles Packer (Letta), or Jesse Vincent (obra) videos and can name specific patterns from them


How to apply (read carefully — this filters applicants)
To be considered, your application must include all four of the following. Applications missing any will be ignored.
1. A link to your GitHub profile. I will look at your repos.
2. A screenshot or short Loom of your own Claude Code, Cursor, or OpenClaw setup showing custom skills, agents, or plugins you have installed and use.
3. Answer this technical question in 3-5 sentences: Why does Anthropic's three-agent harness (planner / generator / evaluator) outperform a single agent with the same model on long-running tasks? What specific failure mode does the evaluator agent prevent?
4. Tell me which of these you have hands-on experience with (be honest, I will verify): LiteLLM proxy, MCP server development, Stagehand or Browser Use, Letta or mem0, n8n production deployment, Qdrant, Telegram bot frameworks, Claude Skills authoring, Agent Teams, Git worktree parallel agent patterns.
Bonus — not required, but will move you to the top of the pile: A 2-minute Loom video walking through one agent or skill you built that you're proud of.

Working style

I am highly technical and will work closely with you. I'll know what you're building and follow along. You won't be working in a vacuum.
We use async-first communication. Daily written updates, weekly video sync.
I move fast and value clear thinking over busywork. If you see a better way to do something, say so.
This role has substantial growth. The starting project is the test. After that, there is 12+ months of ongoing work across multiple businesses — agent fleets, custom skills, MCP server development, autonomous workflows for real estate, video production, and clinic operations.


What happens next
Strong applicants will get a 15-minute video call. We'll discuss the architecture, you'll ask questions, and if it's a fit, we start with a small paid trial task (4–8 hours) before committing to the full starter project.
I'm not looking for someone who needs hand-holding. I'm looking for someone who reads the brief above and thinks "yes, I know exactly how I'd build that, and here are the three questions I'd ask first."
If that's you, I want to hear from you.