Compare two agents in 2 minutes

Start two runs on the same prompt and repository, one with Claude Code and one with Codex, then read them side by side. You need the CLI installed and TRACKS_API_KEY set, see Quickstart with the CLI, and a balance above $0.

Start both runs

PROMPT="Add a tool named get-weather that takes a city and shows its current temperature in a new view. Then run npm run build and fix any error."
REPO=https://github.com/alpic-ai/apps-sdk-template

CLAUDE_RUN=$(tracks runs create "$PROMPT" \
  --harness claude-code --model claude-sonnet-5-5 \
  --repository "$REPO" --setup "npm install" --budget 5 | jq -r .id)

CODEX_RUN=$(tracks runs create "$PROMPT" \
  --harness codex --model gpt-5.5 \
  --repository "$REPO" --setup "npm install" --budget 5 | jq -r .id)

Each command returns right away with the run’s id. Both runs get the same repository, the same setup script and the same $5 budget, so only the agent differs.

Follow them

Watch one agent work as it happens:

tracks runs follow "$CLAUDE_RUN" --events

Or just wait for both to end:

tracks runs follow "$CLAUDE_RUN" > /dev/null &
tracks runs follow "$CODEX_RUN" > /dev/null &
wait

Compare them

for RUN in "$CLAUDE_RUN" "$CODEX_RUN"; do
  tracks runs get "$RUN" | jq '{agent: .agent.harness, end_reason, started_at, ended_at}'
  tracks runs usage "$RUN" | jq '{cost_usd, sandbox_seconds: .sandbox.seconds}'
done

This prints how each run ended, how long it took and what it cost.

Read what each agent did

Open Runs in the web app. Both runs are listed with their agent, cost and duration. Open one to read its conversation: every message the agent wrote and every tool call and command it ran. From the terminal, tracks runs events <id> lists the same events as JSON.

An end reason of completed means the agent stopped, not that it did the job well. Read its last messages to judge its work, or let Tracks judge it with an experiment.

Prefer clicking? In the web app, Runs then New run starts the same run with a repository picked from your library.

Make the comparison fair

One run per agent is an anecdote: the same agent can succeed once and fail the next time. An experiment repeats each run several times, judges every run on yes-or-no criteria you write, and gives each agent a pass rate.

Measure your MCP server or skill: a complete experiment comparing two agents, with and without your tools.
Write evaluations: criteria that judge runs reliably.

Get started

Concepts

Guides

Compare two agents in 2 minutes

Read what each agent did

Make the comparison fair

​Read what each agent did

​Make the comparison fair

Read what each agent did

Make the comparison fair