Execution Modes

ADR 0002: Execution modes (local, remote, hybrid)¶

Status¶

Accepted (2025-12-11); clarified 2026-05-08 by framework ADR-0033

Context¶

Agents/workflows must run in multiple deployment topologies: - thin clients (mobile/web) calling a backend LLM gateway - backend orchestration calling GPU inference fleets - local/dev mode (everything on one machine)

AbstractCore already provides two compatible boundaries: - in-process python API (create_llm(...).generate(...)) - HTTP server boundary (/v1/chat/completions)

Decision¶

AbstractRuntime supports three execution modes:

Local: in-process AbstractCore LLM + local tool execution
Remote: HTTP to AbstractCore server; tools default to passthrough (untrusted)
Hybrid: remote LLM + local tool execution

Remote AbstractCore server mode uses the Core server's own URL, auth token, base URL allowlists, and provider-key override rules. It must not inherit Gateway bearer tokens or browser origin policy implicitly. If Gateway is the caller, Gateway acts as an explicit Core-server client. Gateway-specific env/config is consumed by Gateway and passed to Runtime through explicit run state, effect payloads, client construction arguments, or Runtime-owned environment variables.

Consequences¶

Thin-mode clients can run the workflow logic while delegating inference to a server.
Remote mode supports AbstractCore per-request base_url routing (dynamic endpoint selection).
Tool execution can be gated by trust/sandbox policy outside the router.