Skip to main content
Session run configuration controls the model and execution tools Athena uses in your session. Use Runtime Config to start from a preset or adjust model tuning, file handling, shell execution, and code execution. These settings are available where your deployment permits customer access to agent settings and you can edit the session.

​
Apply a preset

  1. Open the composer’s three-dot More Options menu, then Agent Context, to view Runtime Config.
  2. Under Start from a preset, choose an available customer preset, such as Research Analyst or Coding Agent (sandbox).
  3. Read the confirmation. Applying a preset replaces every field in the current configuration.
  4. Select Apply preset, or Keep current to cancel.
Preset replacement cannot be undone. Review the model, tools, and execution settings after applying it.

​
Tune the model

  1. Open Execution Environment.
  2. Choose Model from the available choices.
  3. Turn on Override temperature if you need an explicit sampling temperature, then adjust the slider.
  4. Set Max output tokens to limit the response output. Leave it blank to use the model maximum.
Model-specific behavior still applies. Models selected with a -thinking suffix force temperature to 1.

​
Choose file and code behavior

  1. Open Filesystem & Sandbox.
  2. Choose Filesystem Mode. Virtual filesystem provides file tools without a shell. Sandbox filesystem uses the session’s sandbox for files and shell execution. Blended interpreter + sandbox lets interpreter code use sandbox file tools and execution.
  3. Use Sandbox (execute) to control sandboxed shell execution.
  4. Turn on Code interpreter (eval) to let Athena execute code.
  5. Choose Interpreter State: Thread (persistent) keeps state across calls and turns; Turn resets it between turns; Call (fresh) starts each call with a blank namespace.
  6. Turn on Programmatic tool calling (PTC) to make enabled tools callable from interpreter code as tools.<name>().

​
Good to know

You cannot change the form while Athena is running or when you have view-only session access. Runtime availability depends on your deployment. Approval-gated tools are excluded from programmatic tool calling. Approval for a whole interpreter run belongs on eval. These settings do not remove workspace approval requirements.