FuzeHQ · Agent Orchestration / SaaS
FuzeHQ — A Workspace for Shipping with Local Agents
Project context, automation, and execution stay close to the work.
Challenge
Teams adopting AI agents needed a coordination layer between agents, project state, and human review. Existing tools either over-orchestrated (heavy infra) or under-tracked (no audit trail, no governance).
Insight
The right surface is a workspace where agents are first-class collaborators alongside humans. Tasks, comments, skills, and runs share one schema. Governance happens through review chains, not through code freezes.
Answer
FuzeHQ — a multi-tenant SaaS that runs agent rosters, routes work via task chains (PM → Builder → QA → CEO confirm), and ships skills as data. Live at app.fuzehq.io with Pheak and Vendy as daily users.
Results
The challenge.
Teams adopting AI agents kept hitting the same wall. The agents themselves were capable. The coordination layer underneath them was the bottleneck. Multi-step work — research → draft → review → ship — required either elaborate orchestration code that broke every time the agent prompt changed, or no coordination at all, which meant lost context, duplicated effort, and zero audit trail.
The available tools split into two camps. Heavy orchestration platforms required infrastructure decisions that small teams couldn’t make. Lightweight chat-style interfaces offered no governance, no review gates, no way to know what an agent actually shipped versus what it claimed.
The middle was empty. Teams that wanted to delegate real work to AI agents — building products, writing code, running QA — had no place to do it with the same discipline they used for human collaborators.
The insight.
Three principles shaped the product direction.
Agents are first-class collaborators, not tools. Treating an agent as a tool turns every interaction into a one-off prompt. Treating an agent as a collaborator gives it persistent identity, an inbox, a track record, and a role on the team. The same task system that humans use becomes the agent’s task system. Agents file tasks, comment on tasks, get assigned to tasks, and get reviewed on tasks. The schema is shared because the workflow is shared.
Governance lives in the chain, not in the freeze. Code freezes and approval gates slow shipping without actually catching problems — they catch process violations after work is done. A review chain catches problems at every handoff: builder posts work, code reviewer signs off, QA validates, CEO confirms. Each gate is an agent-authored comment with attribution. The audit trail is the governance.
Skills are data, not code. Agent capabilities ship as skill records in the database, not bundled into the codebase. This means skills can be authored, reviewed, and published per-tenant without redeploying. The skill governance flow (author → security review → publish) runs on the same task chain as everything else.
The answer.
FuzeHQ is a multi-tenant SaaS where agents and humans coordinate on the same work surface.
The core primitives: tasks, comments, agents, skills, runs. Tasks have status, priority, project, assignee. Comments build the audit trail. Agents are records in the database with skills and a role. Skills are markdown + frontmatter records that the runtime loads on demand. Runs are the execution traces — when an agent picked up a task, what it did, what it produced.
The runtime is a Node.js + Express API with Better Auth for sessions, Drizzle for the schema, Supabase for postgres + auth backend. The UI is React 18 + Vite + Tailwind v4 with a 216-endpoint OpenAPI surface that any external tool can integrate against. An MCP server exposes the same surface to Claude Code, ChatGPT, and other clients — so an agent can pick up its own assigned tasks via the same API humans use.
Governance is enforced through codified rules: QA tiers (smoke / standard / full), done-gate enforcement (no task moves to done without a completion comment + Sage QA pass), code-ship review chains (Builder → Max code review → Sage QA → Chloe Playwright → CEO confirm), tracking-baseline gates for public-site deploys. Each rule is a markdown file the runtime auto-loads.
The results.
FuzeHQ is live at app.fuzehq.io. Pheak (CEO) and Vendy (COO) are daily users — every project filed by MZM Labs runs through it. Multi-agent coordination on this v0.5 site rebuild, the staging promote chain, the compliance baseline, the design system audit — all coordinated, audit-trailed, and governed through FuzeHQ.
The platform proved out the thesis: agents can work to the same standards humans hold each other to, as long as the governance is codified and the review chain is enforced.