Senior AI Automation Engineer — Build a Kanban-Based Multi-Agent Creative Production System

We’re looking for a senior engineer to build a reliable, production-grade AI Creative Machine that coordinates multiple specialized AI agents through a persistent Kanban workflow. This is not a basic chatbot, prompt chain, or Zapier automation. The system must accept a creative mission through one user-facing operator, break it into dependent tasks, route each task to the correct specialist agent, validate every handoff, manage revisions automatically, and stop at explicit human approval gates before paid generation or publishing. The workflow will be used to research customer avatars, analyze competitor advertisements, develop strategy, write long-form direct-response copy, review the copy, create image-generation instructions, review generated assets, and prepare approved packages for external systems. Core Workflow A typical mission should progress automatically through stages such as: Mission intake and validation Competitor-source qualification Customer-avatar matching Creative strategy and mechanism compilation Long-form copywriting Independent copy review Conditional revision routing Visual translation and image-request creation Independent image-request review Human authorization for paid image generation Image generation Pixel-level asset review Human approval External publishing or record creation Each specialist must be a real independently executed agent, not one model pretending to be several agents. What You Will Build 1. Persistent Kanban workflow engine Tasks, dependencies and DAG-based progression Todo, running, blocked, review and completed states Automatic promotion of downstream tasks Bounded retries Conditional repair branches Cancellation and safe resume Mission-level status and progress tracking Historical task and decision audit trail Support for multiple concurrent missions 2. Multi-agent orchestration The system will include specialist agents for functions such as: Mission operator Customer/avatar researcher Creative strategist Direct-response copywriter Independent copy reviewer Copy repair editor Visual strategist Asset reviewer Requirements: Each task runs through the assigned agent/profile Agents receive bounded, versioned assignment packets Inputs and outputs are hash-verified Every execution records the real agent, model, provider and session/run ID A central operator coordinates the workflow without impersonating specialists Failed or invalid work cannot silently move downstream 3. Deterministic validation layer Judgment-heavy work belongs to agents. Deterministic code should handle: Schema validation Required-field validation File and artifact hashing Provenance receipts Task-state transitions Retry limits Idempotency Approval state Cost/spend gates External read-back verification Duplicate detection Structured error reporting 4. Creative quality gates The system must validate more than whether a JSON file exists. Examples include: Correct customer avatar and awareness level Required customer symptoms, desires, fears, objections and proof needs Evidence-backed source quotations Locked market-specific messaging and mechanisms Competitor-ad structure preservation No internal process narration in reader-facing copy No repeated padding No accidental residue from the competitor’s market No fabricated testimonials or unsupported evidence Independent PASS, repairable failure and fundamental failure verdicts Downstream invalidation after fundamental failures 5. Human approval and spending controls The system must fail closed around: Paid AI image generation Cloud-document creation Spreadsheet/database writes Publishing Advertising-platform actions Paid generation and external publishing must have separate explicit approval gates. No advertisement should ever launch without human approval. 6. Slack or similar operator interface We want one friendly user-facing AI operator accessible through Slack. The operator should be able to: Start a mission Report concise progress Explain blockers Request approvals Provide copy and images as files Resume or cancel work Answer status questions Technical logs, hashes and raw JSON should remain in the Kanban audit layer unless requested. 7. External integrations Later stages may connect to: Google Drive Google Sheets Notion or another database AI image-generation providers Competitor-intelligence APIs Meta advertising systems All external writes must be: deterministic; idempotent; explicitly authorized; read back and verified; safe to retry without duplicate records. Technical Expectations We are open to the exact stack, but you should be comfortable with: Python Agent orchestration frameworks or custom orchestration LLM APIs from multiple providers Structured outputs and JSON Schema Persistent workers/controllers DAG or workflow engines Kanban/task-management systems Slack APIs and Socket Mode OAuth integrations Google APIs Notion APIs Image-generation APIs SQLite/PostgreSQL Background services on macOS or Linux Unit and integration testing Experience with Hermes Agent, LangGraph, Temporal, Prefect, Celery, n8n, CrewAI, AutoGen or similar tools is useful, but we care more about sound architecture than a specific framework. Non-Negotiable Engineering Requirements No single-agent role simulation No fake agent attribution No manually advanced “automatic” workflows No success claims based only on mock data No silent fallback when an agent or provider fails No uncontrolled infinite retry loops No credentials committed to the repository No paid calls during development without approval No destructive Git operations No publishing without explicit authorization Every critical state transition must be recoverable and auditable Deliverables Architecture document and workflow diagram Persistent Kanban mission controller Multi-agent routing and execution layer Agent assignment and receipt contracts Schema and semantic validation system Conditional review and repair routing Human approval and spend controls Slack operator interface External-integration abstraction layer Local development and production deployment setup Monitoring, heartbeat and error diagnostics Unit, integration and end-to-end tests Operational documentation Recorded demonstration of a complete mission Clean, maintainable repository with installation instructions Definition of Done The project is complete when we can submit one mission through Slack and observe it automatically progress through real specialist agents and Kanban dependencies until it reaches the correct human approval gate. The demonstration must prove: real specialist-agent execution; automatic handoffs; persistent recovery after restart; input/output provenance; conditional revision routing; no duplicate execution; no unauthorized paid generation; no unauthorized external writes; clear operator updates; complete audit history. Ideal Candidate You have previously built at least one of the following: Multi-agent production workflow Durable AI workflow engine Human-in-the-loop automation platform Agentic content-production pipeline Kanban/DAG orchestration system Reliable LLM system with validators and approval gates You understand that a production agent system is primarily a state, orchestration, provenance and failure-recovery problem, not simply a prompting problem. How to Apply Please begin your proposal with: CREATIVE MACHINE Then answer these questions: What is the most complex multi-agent workflow you have built? How did you prove that individual agents actually executed their assigned work? How would you combine a Kanban interface with a dependency graph? How would you prevent a failed copy review from reaching paid image generation? How would you make external writes idempotent and safely retryable? How would the system recover if the controller, computer or API provider went offline halfway through a mission? Which components would you build deterministically rather than delegate to an LLM? What stack would you recommend, and why? Please share relevant repositories, architecture diagrams or demonstrations. What would you deliver in a paid initial technical trial? Generic proposals or portfolios consisting only of chatbots will not be considered.