Where work runs¶
Agentagon separates the coding agent that reasons about your application from the runner that executes a benchmark. Choose the execution environment when you save a profile, before measuring a baseline.
Your coding host¶
Codex or Claude Code reads permitted evidence, proposes changes, and arranges independent reviews. Candidate authoring stays in local worktrees even when benchmark execution uses a remote runner. Your host's permissions and model provider determine where reasoning occurs.
A code-only audit inspects captured source without running the application. It needs neither an execution profile nor a trace provider. See the quickstart.
Benchmark execution¶
| Runner | Where commands execute | What you arrange |
|---|---|---|
| Local | On the host running the CLI. | Dependencies and appropriate local permissions. |
| SSH | On an existing remote machine. | An authorized destination and permission to transfer the declared execution inputs. |
| E2B | In a configured E2B sandbox. | The optional dependency, a template, and a credential reference. |
Local worktrees protect the original checkout's files; they are not a machine or network sandbox. Remote runners receive frozen execution inputs and return evidence. Only declared private inputs should be transferred.
Use Settings and profiles for your first setup, or the profile and runner reference for exact fields, concurrency requirements, and limits.
A stable environment for comparison¶
A run keeps its frozen execution profile. Editing saved settings affects future runs. Changing the execution environment requires a fresh baseline, so results remain comparable.
Shared hardware and external model services can introduce variation. Inspect repeated measurements and the reported range when judging a candidate.
Saved evidence and the dashboard¶
Workflow evidence lives in the application's ignored .agentagon/ directory. The dashboard reads that checkout's saved projections; it does not run a separate model or reconnect to a runner when you open it.
See Privacy and data for credential handling, model processing, optional Intelligence, and telemetry.