Human-in-the-loop agent orchestration
Run a team of AI coding agents — and stay in control.
cyppie-agents is a workspace where you coordinate specialized agents like teammates: real roles, real permissions, every action visible. You bring the agents; you stay the architect.
Open a workspace, brief a coordinator, watch the work happen.
your workspace
Not an autopilot. A workspace — where you orchestrate a team of specialized tools and keep the controls.
The hype says the AI writes your software while you watch. That's not this. cyppie-agents treats agents as first-class teammates you direct — with a coordinator, defined roles, and a clear chain of who may do what.
Every task, message, grant, and commit is visible and audited. You review the work, you hold the permissions, and you decide what ships. The leverage is real; the control stays yours.
How it works
One coordinator, many workers, you above it all.
Agents are organized hub-and-spoke: a Product Owner coordinates, workers execute in their own isolated checkouts, and nothing routes worker-to-worker. You talk only to the coordinator.
Brief the coordinator
Post a task in the shared channel — “add an endpoint with a test.” One clear ask, marked as a task.
The team executes
The PO decomposes and delegates to workers through their private channels. Each works in its own worktree, commits, and pushes a branch.
You review & ship
Status flows back up to the coordinator, then to you — with real branches in your repo. You merge on your terms.
Built for control
Everything visible. Everything permissioned.
Hub-and-spoke by design
A single coordinator, clear delegation paths, no chaotic agent-to-agent chatter. Structure that stays legible as the team grows.
Human in the loop
You brief, you observe live, you approve. Bring teammates into any channel with a single toggle — read-only, or granted to post.
Permissions first
An access matrix decides who reads and who writes, per channel — and it's fail-closed. No grant, no access. Ever.
Audited by default
Every task, grant, and lifecycle change lands in a metadata-only event log. Nothing important happens off the record.
Isolated worktrees
Each agent works in its own checkout of your repo and pushes its own branch. Parallel work, no stepping on each other.
Your agents, your keys
Built around the Claude agents you already run, on your own credentials. The platform coordinates — it never takes the wheel.
Get started
Spin up your first team in about fifteen minutes.
Open a workspace, add a coordinator and a couple of workers, and hand them a task. The getting-started guide walks you through every step.