Runner
The core component in the OpenAI Agents SDK responsible for executing Agents, offering multiple execution modes (synchronous/asynchronous), connecting to OpenAI servers and returning the final result.
Core Facts
Timeline (last 90 days)
If a critical patch involves permission or token verification logic, symptoms such as Runners going offline, variables failing to decrypt, or Webhook signature verification failures may occur after the upgrade, so validation should cover these sensitive paths.
GitLab's CI/CD variables are stored in the database and can be injected into pipelines; Runners establish a trust relationship with the instance via registration tokens; deployment credentials (Deploy Keys, Deploy Tokens) are used to access external repositories or environments.
LangBot 4.11 changes Runner from built-in code to a plugin installable from the plugin marketplace; the core no longer presets the running logic of service APIs, and all external integrations must be loaded as plugins.
RunConfig can pass a provider object via the model_provider parameter, used at the Runner.run / run_sync call site to replace the global default configuration.
Runner.run_sync and asyncio.run(Runner.run(...)) have exactly the same effect.
The system consists of two components: the control plane is responsible for managing error messages, configuration, and run history; the Runner is the user's own GitHub Actions runner, which holds the code and the AI keys.
A free cloud server only has the user's Shell environment, not the Runner tags, does not hold the project's protected variables, and does not generate the Job Token that scripts expect.
A tautology refers to a case where an assertion cannot fail (such as comparing a value to itself or checking a constant that the test just wrote), which receives zero credit; the classification work must be completed by the Runner.
If the Runner cannot delay publishing logs, that run should be treated as E_LOGS_EARLY and the freeze rejected, rather than pretending the barrier took effect.
The core idea of the review gate is that failure logs must not be readable until after the Runner summary is sealed, and the power to judge must be given to the Runner executing the tests, not the patch author.
2 more timeline events
All Facts (13)
The OpenAI Agents SDK's Runner does not automatically save and maintain conversation history; each call to Runner.run() is independent.
65%UnverifiedIf a critical patch involves permission or token verification logic, symptoms such as Runners going offline, variables failing to decrypt, or Webhook signature verification failures may occur after the upgrade, so validation should cover these sensitive paths.
50%UnverifiedGitLab's CI/CD variables are stored in the database and can be injected into pipelines; Runners establish a trust relationship with the instance via registration tokens; deployment credentials (Deploy Keys, Deploy Tokens) are used to access external repositories or environments.
50%UnverifiedLangBot 4.11 changes Runner from built-in code to a plugin installable from the plugin marketplace; the core no longer presets the running logic of service APIs, and all external integrations must be loaded as plugins.
50%UnverifiedRunConfig can pass a provider object via the model_provider parameter, used at the Runner.run / run_sync call site to replace the global default configuration.
50%UnverifiedRunner.run_sync and asyncio.run(Runner.run(...)) have exactly the same effect.
50%UnverifiedThe system consists of two components: the control plane is responsible for managing error messages, configuration, and run history; the Runner is the user's own GitHub Actions runner, which holds the code and the AI keys.
50%UnverifiedA free cloud server only has the user's Shell environment, not the Runner tags, does not hold the project's protected variables, and does not generate the Job Token that scripts expect.
50%UnverifiedThe core idea of the review gate is that failure logs must not be readable until after the Runner summary is sealed, and the power to judge must be given to the Runner executing the tests, not the patch author.
50%UnverifiedIf the Runner cannot delay publishing logs, that run should be treated as E_LOGS_EARLY and the freeze rejected, rather than pretending the barrier took effect.
50%UnverifiedA tautology refers to a case where an assertion cannot fail (such as comparing a value to itself or checking a constant that the test just wrote), which receives zero credit; the classification work must be completed by the Runner.
50%UnverifiedThe Runner (executor) is responsible for actually invoking the Agent under test to run, rather than simulating its behavior.
50%UnverifiedThe Runner executes only according to the checklist, does not add cases on the fly, and does not modify expected results based on impressions.
50%