The Harness/Runtime × backend routing

Runtime × backend routing

The two orthogonal axes — runtime and compute backend — the resulting matrix, the per-tokenId selection precedence, and why launched agents always route to Hermes on 0g-compute.

A Slopstock agent's behavior is set by two independent choices: which runtime wraps the model call, and which backend the call physically goes to. These are orthogonal — any runtime can sit on any backend — and the router resolves both, per tokenId. The source of truth is the operator's runtime/index.ts.

Two orthogonal axes

The first axis is the runtime — the layer above the LLM call:

• hermes — the full stateful harness: the task loop, tools, skills, and three-layer memory described across the rest of this section.
• openai-compat — a single-shot LLM call wrapped in the same runtime interface. Stateless: no loop, no tools, no skills, no memory.

The second axis is the compute backend — where the call physically goes:

• 0g-compute — routed through 0G's compute SDK for sealed, TeeML-verified inference inside an Intel TDX (or H100/H200) enclave. This is the sealed inference path that yields an attestation.
• openai-compat — plain HTTP to any OpenAI-shaped endpoint (Ollama, OpenRouter, and the like). Unsealed; the attestation slot in the receipt is empty by design.

Because the axes are independent, the runtime kind and the backend kind compose freely — which is exactly what the matrix below enumerates.

The matrix

The cell that matters in production is hermes × 0g-compute: the full harness on sealed inference. The other three cells exist for development and for cheaper modes — the harness on a plain endpoint (stateful but unsealed, good for local iteration), a single sealed call with no agent loop (cheap attested inference), and a single plain call (the simplest baseline).

Selection precedence

For a given tokenId the router resolves the runtime, then the backend, each with an env override that falls back to a global default:

Runtime resolves before backend, and each axis independently prefers its per-token override before falling back to the global default. The chosen backend is held inside the runtime instance. Hermes runtimes are built per-token (state is per-token), each with its own backend; openai-compat runtimes are stateless, so the router caches and shares them by backend kind rather than constructing one per token.

Launched agents

There is one override that sits above the precedence table. Agents minted through /launch live in a dynamic registry, and their config — runtime, model, system prompt — comes from that registry, not from the static env. For every such agent the rule is fixed: launched agents always route to Hermes on 0g-compute, on the single shared 0G compute v4 backend (deepseek-v4-flash on 0G mainnet). The router seeds the agent's system prompt into its data dir and returns a fresh HermesAgentRuntime on the sealed backend — deliberately not cached, because the registry can change at runtime.