We're building Agent Flow — a visual canvas where a node is a live coding agent and an edge is a two-way message bus, with automation triggers and a receipt for every run. It's the one category nobody owns yet. Rather than announce it as "done," we're building it in public and labeling exactly what's real.
Here's where Wave 1 stands.
The engine before the canvas
A visual canvas over agents is only as good as the engine underneath. So Wave 1 isn't UI — it's the fleet control loop and the two-way communication harness. The demo we're building toward:
An orchestrator spawns two implementer agents and a reviewer in parallel. Implementer A finishes and messages the reviewer directly — "diff ready on branch feat/sse." The reviewer reads A's screen, reviews, and messages the orchestrator — "A passes, B still running." The orchestrator reacts to the completion event (no polling) and tears the fleet down.
What's proven
We prototyped the architecture end-to-end, then ported it into the product as tested code. What's landed and green today:
- A backend abstraction ("our own tmux") — one interface that drives a fleet of real agent processes: spawn, send, read-screen, key, close, list, subscribe.
- Two-way comms — any agent can message any agent or the orchestrator. Delivery works both ways: pull (an agent checks its inbox) and push (a message is injected straight into another agent's input). Flat topology; roles organize responsibility, not the communication channel.
- Event-driven completion — the orchestrator reacts to "done" events instead of polling.
- Guardrails — rate limits and de-duplication so a fleet can't spiral into a message storm.
It runs with zero API keys in continuous integration and adds no heavy dependencies — the foundation is deliberately boring and solid.
What's next
- Wave 1 finish: a full PTY/tmux backend, per-tool adapters (Claude Code, Codex, local models), and per-run receipts.
- Wave 2 — the canvas: our own canvas engine (built and maintained in-house, no third-party graph library) where you draw the fleet, plus the live overlay that turns the same graph into a real-time cockpit, plus trigger-driven fleets ("PR opened → spawn a review fleet → post to Slack → human approval → merge").
Honest status
Agent Flow is early access / in active development — it is not a shipped feature yet, and we won't claim otherwise until the demo above runs on real agents end-to-end. When it does, we'll flip the labels and show you the recording.
Want to follow along or get early access? Read the Agent Flow vision →
- The thesis: Visual AI Agent Orchestration: The Missing Category
- Comparisons: vs Zapier/n8n · vs cmux · vs LangGraph Studio · vs CrewAI/AutoGen