25 Agent comparison
Checked 2026-09-12. Product interfaces and policies change. Verify the installed version and official documentation before teaching.
| Harness | Instructions and precedence | Skills/hooks | Approval and isolation |
|---|---|---|---|
| OpenCode | AGENTS.md from project upward; CLAUDE.md fallback; global config and instructions can add files/URLs |
.opencode/skills, compatibility .claude/.agents; plugins include tool.execute.before/after |
opencode --auto is the TUI mode; opencode run --auto "<prompt>" is non-interactive; allow/ask/deny is UX, not a sandbox |
| Antigravity | global GEMINI.md; .agents/rules/ workspace rules with activation modes; no documented automatic AGENTS.md import |
.agents/skills; plugins/hooks have product-specific APIs; no workshop R skill assumed |
--dangerously-skip-permissions approves calls; --sandbox is separate; Deny > Ask > Allow |
| Codex | root→cwd AGENTS.override.md/AGENTS.md; nearer files later; global files also apply |
.agents/skills; hooks can add context, deny, or rewrite supported inputs |
--ask-for-approval never differs from --sandbox danger-full-access; combined bypass removes both |
| Claude Code | CLAUDE.md imported here from AGENTS.md; files layer root→cwd; managed policy can dominate |
nested .claude/skills; PreToolUse hooks can deny/modify |
--dangerously-skip-permissions changes permission mode, not OS isolation; sandbox is separate |
“Auto approval,” “approval bypass,” and “sandbox removal” mean different things. A permission hook can deny or transform a tool request. A container or VM provides the environment boundary. Ask each harness what it loaded, then verify the files, diffs, and command evidence.
25.1 Further reading
- Codex documentation — Check the current OpenAI product rather than transferring settings from OpenCode.
- Claude Code security — Read Claude’s own trust and permission model, including its limits.
- Antigravity permissions — Confirm current precedence and modes for the Google comparison.