Agent Loops 2026: Ralph Loop, /loop, /goal and Headless Loops for AI Coding Agents

Geoffrey Huntley describes Ralph as a technique, and says that in its purest form Ralph is a Bash loop. The loop pipes the same PROMPT.md into the agent on every pass. Work persists in files and git, so each pass sees what the last one did.

while :; do cat PROMPT.md | claude-code ; done

• Ask for one thing per loop. Huntley repeats this rule on purpose.
• His example prompts keep a fix_plan.md as a prioritised to-do list and an AGENT.md with build and run notes.
• He says Ralph can be done with any tool that does not cap tool calls and usage.
• He also describes Ralph as one process working in one repository, doing one task per loop.

Claude Code documents three ways to keep a session running. /loop re-runs a prompt on a time interval. /goal keeps starting new turns until a separate model judges your condition met. A Stop hook in settings decides after every turn whether Claude should continue.

/loop 5m check if the deployment finished and tell me what happened
/loop check whether CI passed and address any review comments
/goal all tests in test/auth pass and the lint step is clean
/goal clear

• /loop is session-scoped. With no interval, Claude picks a delay between one minute and one hour after each iteration.
• A bare /loop runs a built-in maintenance prompt, or your .claude/loop.md or ~/.claude/loop.md if one exists.
• Recurring /loop tasks expire 7 days after creation, and a session can hold up to 50 scheduled tasks.
• Press Esc to stop a self-paced /loop. Set CLAUDE_CODE_DISABLE_CRON=1 to disable the scheduler entirely.
• /goal is a wrapper around a session-scoped prompt-based Stop hook. It does not change your permission mode, so run it in auto mode for unattended turns.
• To bound a goal, add a clause such as "or stop after 20 turns" to the condition.
• A Stop hook that exits with code 2 prevents Claude from stopping and continues the conversation.

• Claude Code hooks
• Hooks across AI coding agents

The anthropics/claude-code repository includes a ralph-wiggum plugin. It runs the loop inside your current session, so you do not need an external Bash loop. A Stop hook blocks Claude's exit and feeds the same prompt back.

/ralph-loop "Build a REST API for todos. Requirements: CRUD operations, input validation, tests. Output <promise>COMPLETE</promise> when done." --completion-promise "COMPLETE" --max-iterations 50

/cancel-ralph

• --max-iterations stops after N iterations. The default is unlimited.
• --completion-promise is a phrase that signals completion. It uses exact string matching.
• The README says to always use --max-iterations as your primary safety mechanism.
• It lists tasks that need human judgment, one-shot operations, unclear success criteria and production debugging as poor fits.

• Claude Code plugins

To script your own loop, run Claude Code non-interactively with -p. It exits with code 0 on success and non-zero on failure, so a shell script can branch on the result. Print mode adds two caps that suit loops: --max-turns and --max-budget-usd.

while :; do
  claude -p "$(cat PROMPT.md)" \
    --max-turns 30 \
    --max-budget-usd 5.00 \
    --permission-mode acceptEdits || break
done

• --max-turns limits agentic turns in print mode and exits with an error when the limit is reached. There is no limit by default.
• --max-budget-usd stops the run once estimated spend reaches the amount. Spend can pass the cap, so leave headroom.
• --output-format json includes total_cost_usd, which a script can add up across passes.
• --continue or --resume carries one conversation across runs. A fresh call per pass gives each pass a clean context.
• Setting a goal with -p runs the /goal loop to completion in one invocation.

• Claude Code headless mode (claude -p)
• Headless mode across AI coding agents
• Claude Code cost and usage limits

Codex has a /goal command for one durable objective. OpenAI says Codex can then work independently for multiple hours and stops when it is confident it has reached the stopping condition. For scripts, codex exec runs Codex non-interactively.

/goal <objective>
/goal pause
/goal resume
/goal clear

codex features enable goals

codex exec --sandbox workspace-write "<task>"
codex exec resume --last "fix the race conditions you found"

• If /goal is missing from the slash command list, enable features.goals in config.toml or run codex features enable goals.
• By default, codex exec runs in a read-only sandbox. Pass --sandbox workspace-write to allow edits.
• --full-auto is no longer supported. Use --sandbox workspace-write instead.
• Use danger-full-access only in a controlled environment such as an isolated CI runner or container.
• codex exec requires a Git repository unless you pass --skip-git-repo-check.

• Codex exec (non-interactive)
• Codex sandbox modes

These agents have no documented built-in loop command, but each has a headless mode you can call from a shell loop.

• Gemini CLI: headless mode starts with -p or --prompt, or in a non-TTY environment. Exit code 53 means the turn limit was exceeded.
• Gemini CLI: model.maxSessionTurns caps turns in a session, and -1 means unlimited. model.disableLoopDetection turns off automatic infinite-loop detection, so leave it off.
• Gemini CLI: YOLO mode, which auto-approves all actions, can only be enabled on the command line with --yolo or --approval-mode=yolo.
• Gemini CLI: for unpaid tier and Google One users, Gemini CLI was replaced by Antigravity CLI on June 18th, 2026.
• Amp: amp -x (--execute) sends a prompt, waits until the agent is done and prints its last message. Execute mode still exists in the rebuilt Amp CLI.
• Amp: after the rebuild, Amp no longer asks for permission before running tools. The old --dangerously-allow-all behavior is the default for users who have not configured permissions.
• Factory: droid exec runs a task without interactive mode. The default autonomy is read-only, --auto low, medium or high raises it, and -s continues a session.
• Factory: --skip-permissions-unsafe skips all permission prompts, though command blocks still apply.

gemini -p "<task>"
amp --execute "<task>"
droid exec --auto medium "run tests"

• Gemini CLI headless mode
• Gemini CLI YOLO mode
• Migrate Gemini CLI to Antigravity CLI
• Amp settings
• Factory Droid config

Cursor runs long jobs in the cloud instead of in a local loop. Its help docs describe background agents, also called Cloud Agents, as coding agents that typically run for a very long time.

• Each Cloud Agent runs on its own dedicated virtual machine, sandboxed and isolated from your local machine.
• You control which repositories, secrets and network access an environment gets.
• In February 2026 Cursor opened its long-running agents research preview at cursor.com/agents for Ultra, Teams and Enterprise users.
• Long-running agents propose a plan and wait for approval before they start executing.

• Cursor sandbox
• Cursor CLI

• Give every loop a hard stop: --max-iterations for the Ralph plugin, --max-turns for claude -p, or a turn clause in a /goal condition.
• Cap spend with --max-budget-usd on claude -p, and track total_cost_usd from JSON output.
• Run in a sandbox. codex exec defaults to read-only, and Cursor Cloud Agents run in isolated VMs.
• Keep broad flags such as --dangerously-skip-permissions, danger-full-access, --yolo and --skip-permissions-unsafe for isolated runners or containers.
• Write a completion condition the agent can prove in its own output, such as a test command exiting 0.
• Remember that /goal can stop on its own: it clears on an exhausted credit balance or an authentication failure.

• Sandbox AI coding agents
• Claude Code sandbox
• Allow all commands in AI coding agents
• Claude Code permissions