Flare Control

Flare Control is an agent orchestration platform built entirely on Cloudflare’s edge. It lets you create AI agents, give them tools and memory, run them on a schedule or on demand, and keep a human in the loop for anything sensitive — all without running a single long-lived server.

At a glance

• Role: Sole engineer, product architecture through implementation.
• Scope: Agent control plane, queue runtime, dashboard, approvals, memory, scheduled runs, and Telegram integration.
• Stack: Cloudflare Workers, Hono, Queues, D1, R2, Vectorize, Durable Workflows, Durable Objects, React, Vite, TypeScript.
• Status: Working reference platform and public case study.
• Proof: Near-zero idle cost, queue-driven execution, durable approval flows, and full run auditability.

The problem

Most agent frameworks assume an always-on backend: a process that holds state, polls queues, and runs the loop. That’s expensive to operate, hard to scale, and gives you little visibility when an agent does something unexpected. Teams that want autonomous agents in production also need the opposite of “fire and forget”: audit trails, spend limits, and an approval gate before an agent takes a consequential action.

What I built

I designed and built the platform end-to-end as the sole engineer:

• A control-plane API (Hono on Workers) for managing agents, sessions, tools, and enqueuing runs.
• A queue-driven runtime that executes agent tasks off Cloudflare Queues, so there’s no process to keep warm and work scales horizontally on its own.
• Durable Workflows that coordinate long-running, human-in-the-loop sequences — an agent can pause mid-run, wait for an approval, then resume exactly where it left off.
• A React + Vite dashboard to create agents, inspect runs, replay or cancel them, and manage scheduled tasks with pause/resume/retry.
• Telegram integration for chatting with agents from outside the dashboard.

How it works

The architecture deliberately splits stateless from stateful work. The control plane and runtime are plain Workers; D1 holds relational state (agents, runs, schedules), R2 stores artifacts, and Vectorize powers vector-indexed memory so agents can recall prior context. Models are reached through the Cloudflare AI Gateway behind a provider abstraction, with per-agent model selection. Durable Objects are used narrowly — only where strong consistency is actually required, like session and lock coordination — instead of holding the whole system’s state.

Workspace-level policies cap what each agent can do: which tools and domains it may reach and how much it may spend. Every run is traced, so a run can be opened up after the fact and replayed step by step.

Outcome

The result is a production-grade agent platform with near-zero idle cost: nothing runs until there’s work in the queue, long-running approvals survive restarts via Workflows, and every action is auditable. It doubles as a reference architecture for building serverless, human-in-the-loop agent systems on Cloudflare.