Beyond the Chatbox: LLM Coding Agents for Research

Egor Kotov

5 Other agents

Author

Affiliations

Egor Kotov

TU Dortmund University

Universitat Pompeu Fabra

Max Planck Institute for Demographic Research

The exercises use OpenCode. You may compare another agent if it is already available. This page was checked 2026-09-14, so verify commands in the current vendor documentation.

5.1 Optional comparisons

Antigravity readiness

Open a terminal in the lab directory and run agy. The Codespace already includes the CLI.
If prompted, open the printed authorization URL in your browser, sign in, and return the one-time code to the terminal. This is Antigravity’s documented remote sign-in flow.
Run /model and choose an available model. Ask What directory are you working in? Do not change files.
Before edits, inspect /permissions, the loaded project rules, sandbox setting, and telemetry setting in the official Antigravity documentation. Turning off telemetry does not make research data safe to share.

The Antigravity setup recording may show older screens.

Codex readiness

In the lab directory, run codex login --device-auth and complete the one-time browser flow. This is the documented option for a remote or headless environment such as Codespaces.
Run codex login status; it should report an active sign-in method without displaying a credential.
Run codex, check the model, approval policy, and sandbox shown by the interface, then ask What directory are you working in? Do not change files.

Use the official Codex authentication guide and agent approvals and security guide. Check the account’s current data controls before using anything beyond toy data.

Claude Code readiness

In the lab directory, run claude auth login and follow the documented browser flow. Claude Code requires an eligible Claude plan, Console account, or supported cloud provider.
Run claude auth status. If startup or configuration fails, run claude doctor from the shell.
Run claude, inspect /status, and ask What directory are you working in? Do not change files.

Use Claude Code’s current quickstart and authentication documentation. Claude reads CLAUDE.md according to its own precedence rules; that file does not create an operating-system security boundary.

5.2 Further reading

Codex agent approvals and security — Distinguishes approval policy and sandboxing for Codex rather than borrowing another product’s vocabulary.
Claude Code security — Explains Claude Code’s trust model and the limits of permission prompts.
Antigravity permissions — Use the current rule precedence and wildcard semantics when comparing Antigravity with the supported path.

Approval bypass and sandbox settings are separate controls. Antigravity’s --dangerously-skip-permissions approves all tool calls and differs from --sandbox. Codex’s --ask-for-approval never differs from --sandbox danger-full-access. Claude’s --dangerously-skip-permissions is a permission mode, not a sandbox. Verify current help, and never use these bypass modes on your laptop or with real data.

For every agent, check the working directory, loaded instructions, permissions, and changed files. If it does not run in the Codespace, inspect its documentation or interface and continue with OpenCode.