Terminals, structured agents and review.

1. Run Codex and Claude Code in separate terminals

Use this when you want the agents' normal terminal interfaces and existing authentication. Give concurrent writers separate checkouts. Keep a dev server or test command in another session so it remains visible while you switch agents.

1. Open two Terminal sessions. In the first, change to the checkout assigned to Codex and start codex. In the second, use the checkout assigned to Claude Code and start claude. Authenticate through each tool's normal flow.
2. Give each agent a bounded task. For example, assign the dialog behavior to one agent and ask the other to review a separate module. State the expected result and relevant tests. A separate session alone does not isolate file edits.
3. Follow the actual work. Switch sessions to read prompts and output, answer questions and inspect Git changes. Processes stay on the host when the browser disconnects, while the host and service remain running.
4. Review before merging. Inspect the diff, run the relevant tests and open the app. An agent's completion message does not prove the change works.

Example host commands, run in the respective assigned checkout:

# Terminal A
codex

# Terminal B
claude

# Review the checkout before accepting changes
git status --short
git diff --stat
git diff

See parallel work and checkout ownership and terminal-agent setup. The recording illustrates the interface; it does not benchmark agent speed or establish compatibility for every provider account.

2. Inspect a structured agent, then verify the app

Use Agent Controller when you need conversation entries, tool events and supported approvals instead of a terminal transcript. Codex app-server, Claude Agent SDK, OpenCode HTTP and ACP agents connect through their respective integrations.

1. Select Agent Controller at launch. Choose an installed supported runtime and the intended repository. Check the protocol and prerequisite table for your agent.
2. Describe an observable result. For a UI change, name the interaction and expected behavior. Read the tool output and code changes. Answer approval requests only where the integration exposes them.
3. Run the app in a separate Terminal session. Use the project's own development command. Open its localhost address in Dev Browser beside the conversation and reproduce the interaction.
4. Check both code and behavior. Review the diff and test results, try keyboard navigation and inspect the relevant narrow layout. Close completed sessions after accepting or rejecting the changes.