Visualize and compare runs in AgentKernelArena

Visualize and compare runs in AgentKernelArena#

2026-07-15

2 min read time

Applies to Linux

AgentKernelArena ships the src.visualization module for comparing experiment run reports. Reports identify the agent, target GPU, and run timestamp; model/provider provenance is not currently a dashboard field. This topic covers how to build the dashboard data and serve it.

What the dashboard reads#

The dashboard scans for report directories that contain:

  • overall_summary.csv

  • task_type_breakdown.json

  • overall_report.txt

By default, it scans only visualization-specific report bundles:

.visualization/reports/<report_name>/

Workspace-run reports, which are usually located at workspace_<gpu>_<agent>/run_<timestamp>/reports/, can also be scanned, but this is opt-in.

Build the dashboard data and serve it#

After a normal AgentKernelArena run, reports land in workspace_<gpu>_<agent>/run_<timestamp>/reports/. Pass --include-workspace-runs so the module picks them up. Run the following from the AgentKernelArena repository root:

python3 -m src.visualization build --include-workspace-runs
python3 -m src.visualization serve --host 127.0.0.1 --port 8080

Then open:

http://127.0.0.1:8080

Without this flag, the module only scans .visualization/reports/, which is empty by default, and the dashboard shows no data.

To build and serve with one command, run:

python3 -m src.visualization run \
  --include-workspace-runs \
  --host 127.0.0.1 \
  --port 8080

The equivalent convenience target is make visualization-run.

Serve on port 80#

Port 80 usually requires elevated privileges:

sudo python3 -m src.visualization serve --host 0.0.0.0 --port 80

Rebuild after new runs#

Whenever a run produces new reports, rebuild the dashboard payload with the same flag used at first build:

python3 -m src.visualization build --include-workspace-runs
# or: make visualization-build

The dashboard picks up newly discovered report directories on the next refresh.

If you placed report bundles manually in .visualization/reports/<report_name>/ instead, omit the flag:

python3 -m src.visualization build

Dashboard implementation notes#

The following details apply to the dashboard implementation:

  • Python code and static frontend assets live under src/visualization/.

  • Generated data.js and data.json files are written to .visualization/dashboard/, outside the source tree.

  • Source-file links are exposed through /artifacts/... and /reports/... routes that only allow .csv, .json, and .txt files.

  • If no reports are found yet, the dashboard still builds and shows an empty state.

  • Directory listings, hidden paths, and path traversal are rejected.

Contents