Develop mode

For designing agents, writing skills, training experts, running evaluations, integrating via API, and deploying on your own cluster. Flip to Work mode in the navbar if you only need to chat with deployed agents and watch sessions.

How the platform is built

Surogate is built on open-source components you can run and inspect yourself:

Repo	Role	License
invergent-ai/surogates	Agent runtime — harness loop, sessions, sandbox, MCP, governance, channels, memory, tools	AGPL-3.0
invergent-ai/surogate	Training framework — the SFT/RL engine behind the expert models you own	Apache-2.0
invergent-ai/surogate-hub	Artifact versioning store — “Git for data” for skills, datasets, and models	Apache-2.0

When you deploy an agent at <slug>.cloud.surogate.ai, the platform wraps the runtime image and binds your config to it. The runtime is what’s actually running inside the container.

This documentation focuses on the platform surface — what you can do through ops.surogate.ai. For runtime internals, see the runtime architecture reference.

The builder loop

       ┌─────────────────────────────────────┐
       │                                     │
       ▼                                     │
┌──────────┐    ┌──────────┐    ┌──────────┐ │   ┌──────────┐
│  BUILD   │ →  │ OBSERVE  │ →  │  TRAIN   │ → │  REDEPLOY  │
│ (config) │    │(sessions)│    │ (dataset │ │   │ (new agent│
│          │    │ & evals) │    │ + train) │ │   │  version) │
└──────────┘    └──────────┘    └──────────┘ │   └─────┬────┘
                                              │         │
                                              └─────────┘

Most iteration lives in build → observe. Training experts and re-deploying happens weekly to monthly, not daily.

Start at The builder loop overview for the full cycle, or jump to Use cases for end-to-end recipes.