Private build — Agent as a Service

AGENT COMPANY

An autonomous AI company that drafts, and a human that decides. A full agent workforce running on a single Mac mini.

Project Overview

Agent Company is a self-hosted multi-agent AI company that runs a real business's back office. It's organized like a company: research, content, marketing, brainstorm, and dev offices, each staffed by role-anchored agents built on CrewAI and served by local models, with lead seats routed through a proxy to free cloud models and a guaranteed local fallback. Work flows like this — a request comes in over Telegram, the Planner routes it to the right office, agents pull grounding from a research-brief knowledge base, and the office produces a draft. Nothing executes on its own: every output is a proposal that a human approves before it ships. The result is an agent workforce with the leverage of automation and the safety of human sign-off, currently producing grounded marketing and content for a live education business.

Offered as a Service

Agent Company is offered as a service, built around your company's needs. I tailor the offices, agents, and guardrails to your business and fit the system into the workflow and tools your team already uses, so nothing about how you work has to change while the agents take on the drafting.

Key Features
  • Company-style agent org: research, content, marketing, brainstorm, and dev offices with role-anchored prompts and a universal topic gate
  • Draft-only governance — agents suggest, a human approves and executes via Telegram; nothing ships autonomously
  • Hybrid model routing: free cloud models via proxy with a guaranteed local fallback; the Planner is hard-pinned to local for routing safety
  • RAG-grounded output from research briefs, with stale embeddings purged before every run so old data never bleeds into new work
  • Hard-coded brand constraints (voice, pricing, format rules) and per-workflow token budgets with breach alerts
  • Pixel-art office dashboard for live monitoring, plus post-commit regression review against rules learned from real failure patterns

© 2025 Werner's Works. All rights reserved.