examples

Examples

Two orthogonal features, two deployment modes — four quadrants.

	Local backend (in-process proxy)	Daemon backend (agentopt serve)
Offline model selection	selection/local/	selection/daemon/
Online per-call routing	routing/local/	routing/daemon/

The user-facing API is identical between local and daemon mode — setting AGENTOPT_GATEWAY_URL is the entire switch. See docs/api/router.md and docs/api/proxy.md.

shared/ holds adapters reused by more than one example (e.g. the OpenClawAgent wrapper used by both selection and routing).

Selection vs. routing — when to use which

• Selection picks one combination of models and evaluates it over a whole dataset. The decision is made once, before the run.
• Routing picks a model per individual LLM call at request time, based on prompt content, session history, or a fixed policy.

They're orthogonal and compose: a ModelSelector could evaluate combinations that each include a router.

Prerequisites

Most examples need:

pip install "agentopt-py[examples]"
export OPENAI_API_KEY=...   # or the relevant provider key

Daemon-mode examples need a daemon running in a separate terminal:

agentopt serve                            # default port 9000
# in another shell:
export AGENTOPT_GATEWAY_URL=http://127.0.0.1:9000