Industry

Meta Prepares "Agent Engine" for Meta API Platform: Internal Codename "Forge" Explored

Details surface regarding Meta’s upcoming enterprise developer offering codenamed "Forge," an Agent Engine slated for the Meta API Platform designed to host, orchestrate, and interconnect autonomous enterprise agent swarms built on Llama foundation models.

By Julian Thorne · 2026-10-03 · 12 min read

Internal architectural disclosures reveal that Meta is building an enterprise-grade agent orchestration runtime dubbed "Agent Engine," developed under the internal project name "Forge."

Intended as a major expansion of the Meta API Platform, Forge provides managed hosting, state persistence, and inter-agent communication protocols for businesses deploying autonomous customer-facing and back-office agents.

While Meta previously focused on providing raw open-weights models through the Llama family, Forge signals a commercial pivot toward offering complete managed execution infrastructure.

Architectural Layout: The Forge Agent Stack

Leaked documentation illustrates a three-layer execution framework designed to host persistent agent graphs at high scale:

┌────────────────────────────────────────────────────────┐
│                   Client Touchpoints                   │
│   WhatsApp Business │ Instagram DMs │ Web SDK │ Zapier │
└──────────────────────────┬─────────────────────────────┘
                           │
                           ▼ (WebSocket / Webhook Ingestion)
┌────────────────────────────────────────────────────────┐
│             Meta Agent Engine (Forge Core)             │
│  • Session State Graph        • Memory & Vector Store  │
│  • Agent Swarm Router         • Policy & Safety Shield │
└──────────────────────────┬─────────────────────────────┘
                           │
           ┌───────────────┼───────────────┐
           ▼               ▼               ▼
   [Llama-3.3 70B]   [Llama-4 Scout]   [Micro-Tool Worker]
   Reasoning Core     Router Agent      gVisor Sandbox
  1. Session State Graph: Replaces stateless API calls with long-lived session state machines that remember customer history across weeks of intermittent messaging.
  2. Sub-Agent Triage: Instead of forcing a single model to handle order status, technical troubleshooting, and payments, Forge routes intents across specialized worker sub-agents.
  3. Managed Tool Sandboxes: Executes Python data queries and CRM mutations inside hardened gVisor microVMs without exposing host systems.

Why Meta is Building Managed Agent Infrastructure

Historically, enterprises adopted Meta’s open Llama models by self-hosting them on cloud platforms like AWS Bedrock, Microsoft Azure, or GCP. This dynamic allowed cloud hyperscalers to capture the compute margin while Meta absorbed the multi-billion-dollar pre-training costs.

Forge allows Meta to monetize its open-source ecosystem directly:

  • Zero-Friction WhatsApp & Messenger Deployment: Connecting an agent to billions of active messaging users requires one-click token authorization rather than complex custom webhook bridges.
  • Pay-Per-Active-Agent Pricing: Provides predictable tier-based hosting fees alongside token throughput billing.
  • Enterprise SLA Guarantees: Guarantees low latency and high availability for mission-critical business processes.

Release Horizon and Developer Preview

While Meta has not officially announced a public launch date, internal roadmaps suggest a private developer beta of Agent Engine will launch ahead of Meta Connect, targeting enterprise partners running high-volume customer service workflows.