Model calls
Use supported L3 provider adapters and validate responses before they influence execution.
A provider call from L3 is part of strategy execution. It is separate from Agent Lab's authoring conversation and its connection settings.
Select the runtime adapter
L3 configuration selects the provider and optional model. The current registry includes:
| Provider | Runtime purpose |
|---|---|
fake | Deterministic supplied output for tests and examples |
openai | The repository's Responses-based API adapter |
openai-compatible | Compatible endpoint adapter with supported tool-calling modes |
codex-cli | Local CLI adapter, restricted to administrators |
codex-cli-loop | Local CLI adapter with the supported iterative tool protocol |
Aliases exist in the registry, but use one explicit provider name in a saved example. An unrecognized provider reports that it is unavailable when called.
Call explicitly
Use await provider.provider.complete(request). The response includes provider identity, optional model, output, tool results, and a creation timestamp. output is not guaranteed to be an object with your desired fields; it can be parsed JSON or plain text depending on the adapter's response.
Define and validate the shape your code accepts. For permission decisions, check an explicit Boolean field rather than coercing an arbitrary object or string to true. The fake-provider example demonstrates validation without executing actions.
Access and credentials
Real provider calls are blocked during backtesting and historical warmup by server-owned access rules. Local Codex adapters additionally require administrator access. Changing script state or writing an access flag into request input does not grant that authority.
API adapters require their supported credential configuration; the OpenAI adapter does not automatically share a server environment credential with every user. Agent Lab's saved authoring credential is not automatically the L3 strategy credential.
Treat any credential placed in source/configuration as part of that saved material. Keep secrets out of prompts, logs, examples, and shared revisions. This guide does not introduce a new secret-storage mechanism.
Bound the work
Choose cadence, bounded input, allowed tools, and adapter-supported timeout settings. Supported loop adapters expose a positive maxTurns and an allowed-tool list; other adapters need not implement the same loop behavior.
Provider output does not place an order on its own. Your script decides whether to turn validated evidence into a request. Continue with failures and timeouts.