AI Agent Infrastructure Consultant — Parallel Agents Optimization (Python/Node)

**Overview** We are a fast-growing SaaS company with a lean engineering team (~10 devs) utilizing a modern Python (FastAPI/Django) and Node.js backend, React frontend, and PostgreSQL stack. We have already deployed an initial multi-model agent stack—Claude Code, LiteLLM gateway, Git worktrees, and MCP integrations. We need an expert to run an intensive architecture review and optimization session for our current infrastructure. We are not looking for someone to build a full-time, weeks-long project from scratch. Instead, we need a seasoned engineer who has shipped this exact type of infrastructure end-to-end to audit our setup, identify architectural gaps, and guide our team on hardened implementation. This project must move fast. If your timeline is measured in weeks, please do not apply. We want someone who looks at this scope, jumps into a review session, and delivers actionable architectural guidance in days. This starts as a focused, urgent consultation. However, we expect ongoing advisory work—follow-ups, architecture adjustments, and enhancement reviews—as the AI tooling landscape shifts. For the right engineer, this will turn into a recurring relationship. We are completely open to a fixed price per milestone or an hourly structure. **What You Need to Have Actually Shipped and Can Review (Not Just Read About)** * **Full Agentic Coding Harnesses:** The entire loop: orchestrator → subagent → CI gate → merge loops. * **Isolation Layers:** Configured execution layers (such as E2B, Modal, or secure Docker runtimes) as isolated sandboxes for AI-generated code. * **Parallel Claude Code Sessions:** Managed multiple simultaneous subagents on scoped tasks via Git worktrees. * **Self-Hosted LiteLLM Gateways:** Routing to multiple models (Claude, GPT, Gemini, DeepSeek). * **MCP Server Infrastructure:** Connected file system, PostgreSQL, Atlassian, and Slack tool layers for active agents. * **Agent Framework Structures:** Used CLAUDE.md, COMMON\_MISTAKES.md, subagent role definitions, hook scripts, and settings.json. * **Human-in-the-Loop Orchestration:** Built Plan Mode or equivalent approval gates before agent execution. * **Multi-Agent Frameworks:** 7-agent feature factory patterns or frameworks like LangGraph, CrewAI, or Autogen. * **Durable Workflow Engines:** Applied Temporal, n8n, or similar tools for long-running agent workflow execution. * **Mechanical Quality Gates:** Treating CI green as the ultimate gate for agent output quality. \[[1](https://manveerc.substack.com/p/ai-agent-sandboxing-guide)\] **Our Current Stack (What you are reviewing)** * **Backend:** Python (FastAPI / Django) & Node.js (TypeScript) * **Frontend:** React (Next.js) * **Database & ORM:** PostgreSQL / Prisma / SQLAlchemy * **Infrastructure:** Docker Compose, AWS (ECS/EKS) * **CI/CD:** GitHub Actions / GitLab CI * **AI Layer:** Claude Code with shared `.claude/` directory, CLAUDE.md, and LiteLLM gateway in Docker * **MCP:** Atlassian (Jira/Confluence), custom PostgreSQL MCP server, Slack * **Workflow Automation:** Temporal / n8n * **QA Automation:** Playwright / Autonoma **Scope of Work (Review & Advisory Only)** 1. **Comprehensive Audit:** Audit our current agent harness and identify architectural gaps against a production-grade standard. 2. **Sandbox Strategy Consultation:** Review our environment strategy to ensure highly secure, isolated execution runtimes for agent code runs. 3. **Workflow Hardening Review:** Evaluate our parallel agent workflow setup (Git worktrees, subagent role configs, hook scripts, and settings lockdown). 4. **CI Pipeline Integration Strategy:** Advise on wiring our sandbox execution layer into the existing CI pipeline so agent-executed code runs in clean snapshots, not live infra. 5. **Architectural Runbook:** Deliver an optimization report / documented standard that our backend lead can easily own and execute going forward. **How to Apply** Skip the generic pitch. Show us something real to be considered: 1. A GitHub repo, architecture diagram, or Loom walkthrough of an agentic harness you have actually shipped. 2. Specific tools from our stack you have personally configured (E2B, LiteLLM, Claude Code, etc.). 3. One sentence explaining the hardest problem you solved to get full agent loops running reliably. 4. Your availability to conduct this high-impact architectural review session this week.