Edge Runtime

AgentDeploy edge agents run on Cloudflare Workers with Durable Objects — a serverless execution environment deployed to 300+ cities worldwide.

How SDK-based agents are deployed

When a buyer deploys an SDK-based agent, the platform:

1. Receives the bundled ESM module from ad deploy (already bundled by esbuild)
2. Analyzes the script with analyzeRenderedScript() to detect:
   • Durable Object classes (extends Agent, AIChatAgent)
   • Agents SDK imports (agents, @cloudflare/ai-chat, @agentdeploy/edge-sdk)
   • Workers AI binding usage (env.AI)
   • Agent routing patterns (routeAgentRequest)
3. Generates deployment metadata including:
   • DO namespace bindings for each detected class
   • new_sqlite_classes migration tag for SQLite-backed DOs
   • nodejs_compat compatibility flag
   • Workers AI binding (if env.AI detected)
4. Wraps with AD_ constants* — Injects AD_DEPLOYMENT_ID, AD_MODEL, AD_GATEWAY_BASE_URL, etc. as ambient globals
5. Deploys to Cloudflare via the Workers API with full metadata
6. Binds buyer secrets via the Cloudflare Secrets API
7. Returns a global URL that routes to the worker

The rendered worker

// AgentDeploy Edge Worker — v2.0.0-edge
// Generated from template 42 version 1.0.0

// Platform-injected constants
const AD_DEPLOYMENT_ID = "uuid-here";
const AD_USER_ID = "buyer-uuid";
const AD_TEMPLATE_ID = 42;
const AD_MODEL = "openai/gpt-4o-mini";
const AD_REGION = "auto";
const AD_GATEWAY_BASE_URL = "https://gateway.agentdeploy.io";

// ─── Seller's bundled code (single ESM module) ───
// This is the output of: ad deploy (esbuild)
// Contains: @agentdeploy/edge-sdk + agents + @cloudflare/ai-chat + seller code
export class MyAgent extends AgentDeployChatAgent { ... }
export default createHandler(MyAgent);

Durable Object auto-detection

The platform's analyzeRenderedScript() uses regex patterns to detect DO classes:

// Detection patterns:
/export\s+class\s+(\w+)\s+extends\s+(?:Agent|AIChatAgent|AgentDeployAgent|AgentDeployChatAgent)\b/g
/export\s+(?:const|let)\s+(\w+)\s*=\s*(?:createAgent|createChatAgent)\s*\(/g

Each detected class results in:

1. A durable_object_namespace binding in the Worker metadata
2. A new_sqlite_classes entry in the migration tag (for SQLite-backed DOs)

No manual configuration needed — the platform infers everything from the bundled code.

Worker naming convention

edge-{template_id}-{user_id_8chars}-{timestamp}

Example: edge-42-abc12345-1721640000000

Lowercase, sanitized, max 52 characters (Cloudflare limit).

Execution limits

Resource	Limit
CPU time	50ms default (configurable up to 30,000ms)
Memory	128 MB
Script size (raw)	256 KB
Script size (SDK-bundled)	3 MB
Request body	100 MB
Subrequests	50 per request (1,000 with nodejs_compat)
Concurrent connections	No limit

Compatibility flags

SDK-based agents automatically get nodejs_compat injected when Agents SDK imports are detected. For raw workers, specify manually:

{
  "compatibility_date": "2025-01-01",
  "compatibility_flags": ["nodejs_compat"]
}

Bindings

The platform auto-generates these binding types based on script analysis:

Binding type	Trigger	Use case
durable_object_namespace	DO class detected	Stateful agent storage
ai	env.AI reference detected	Workers AI models
secret_text	Buyer-configured secrets	API keys, tokens
plain_text	AD_* constants	Platform metadata

For KV namespaces, R2 buckets, D1 databases, and queues — declare them in your wrangler.jsonc for local dev, and the platform passes them through if included in the edge version metadata.

Gateway integration

All LLM calls route through the AgentDeploy gateway:

Agent code → useGateway() → AD_GATEWAY_BASE_URL/v1/chat/completions → Gateway
                                                                         │
                                                    ┌────────────────────┤
                                                    │                    │
                                              Token metering    Provider routing
                                                    │                    │
                                              Billing system    OpenAI / Anthropic
                                                                   / Google / Groq

The gateway:

• Uses the X-AD-Deployment header for authentication (no API key needed in agent code)
• Tracks token usage per deployment
• Enforces buyer plan limits (overage protection)
• Writes revshare ledger entries for token markup

Local development

ad dev

This wraps wrangler dev with your project's wrangler.jsonc. The wrangler config includes:

• DO bindings for local testing
• SQLite migrations
• nodejs_compat flag
• Local AD_* variable stubs

The local gateway stubs route to the platform's gateway URL, so billing and metering still work in dev mode.