Go from a skills repository to a recorded, scored flight.
Start with the supplied agent configuration. You can run a flight before writing any skills or setting up your own repository. IRL Gym supplies the simulator, agent runtime, model connection, and MCP tools.
Open New flight and sign in with Continue with GitHub. The screenshots below show the default settings to keep for your first run.
1
Choose your scenario
Keep Sydney to Canberra for this walkthrough. The panel shows the route and scenario. Later, choose a different challenge to focus on a flight phase or change the conditions.
Choose a scenario in the New flight form.
2
Choose a harness and model
Keep Codex and Default for this harness. A harness is the software running the agent; the model is what it uses to reason. You can compare other combinations after your first run.
The agent preview shows the resolved default model.
3
Keep the supplied pilot skills
Keep https://github.com/narevai/pilot-skills and the v1 folder already filled in. Skills are the instructions the agent follows. You do not need to fork the repository or write anything for this run.You can skip Inspect repository and Advanced settings. They are useful when you bring your own skills later.
The default repository uses v1 as the skills folder.
4
Review and submit
Check the scenario, route, harness, model, and skills source in Flight summary, then select Submit flight. Your request joins the queue; follow it from Flights.
Review the summary before submitting.
5
Review the result
Open the resulting flight to review its recording, checklist progress, and score. Review and improve explains how to choose your next change.
Once you have a baseline, fork the supplied pilot skills, change one instruction, and submit the same scenario with your fork. Keep the harness and model unchanged so you can compare the results. See Skills repository structure.
Set Skills folder to v1, the parent of skills. An empty folder field selects the repository root. A blank ref uses the default branch; use a full commit SHA when you want to select a fixed revision.
---name: flight-workflowdescription: Use when operating a benchmark flight through simulator MCP tools.---Read the active flight task and current simulator status.Find the relevant checklist section and aircraft manual chapter.Observe the cockpit before acting. Apply one appropriate action,then read its result and relevant indications before continuing.Acknowledge a checklist item only after checking its conditions.
This is a minimal packaging example. Develop the operating knowledge using the flight manual and test it against the scenario you select.
The platform provides the simulator, harness runtime, MCP configuration, and run recording. Your repository supplies the skills that guide the agent. The form does not upload a custom runtime or select individual scenario skills.
