How I use Claude Code's -p flag to run autonomous AI sessions without any human intervention

#ai #claudeai #automation #devtools

The --print (-p) flag in Claude Code is the most underrated feature for building autonomous AI agents.

Most people use Claude Code interactively. But with -p, you can run fully headless sessions where Claude:

Reads files, writes code, executes commands
Has access to the full tool suite (Bash, Read, Write, Edit, Glob, Grep)
Works within a budget cap (--max-budget-usd)
Outputs its entire work log to stdout

Here's exactly how I use it to run a developer tools business automatically.

The basic pattern

claude \
  -p "You are an autonomous agent. Execute these tasks: ..." \
  --dangerously-skip-permissions \
  --add-dir /path/to/project \
  --model sonnet \
  --max-budget-usd 3

Key flags:

-p "$PROMPT" — non-interactive mode, executes prompt and exits
--dangerously-skip-permissions — bypasses all tool permission prompts
--add-dir — makes additional directories available as context
--max-budget-usd — hard spend cap per session
--model — which Claude model to use

Running it on a schedule

I use macOS launchd plists to trigger sessions 3x/day:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" ...>
<plist version="1.0">
<dict>
    <key>Label</key>
    <string>com.myagent.morning</string>

    <key>ProgramArguments</key>
    <array>
        <string>/bin/bash</string>
        <string>/path/to/agent_wake.sh</string>
        <string>morning</string>
    </array>

    <key>StartCalendarInterval</key>
    <dict>
        <key>Hour</key>
        <integer>6</integer>
        <key>Minute</key>
        <integer>0</integer>
    </dict>
</dict>
</plist>

Load it: launchctl load ~/Library/LaunchAgents/com.myagent.morning.plist

Writing effective autonomous prompts

The prompt structure matters a lot for headless sessions:

You are [agent name] — an AI agent autonomously running [business].
[Human] is NOT present. Execute all of the following without asking any questions.

CONTEXT:
- [Key facts about the business/environment]
- [Available scripts and their locations]
- [Current status/state]

TASKS (execute in order):
1. [Specific task with exact file paths]
2. [Specific task with exact commands]
3. [Write a report documenting what you did]

Work autonomously. If something is blocked, log it and move on. Never stop to ask a question.

Critical rules for autonomous prompts:

Be specific about file paths — no "somewhere in the project"
Give fallback instructions — "if X fails, log it and do Y instead"
End with a report task — the agent documents its own work
Set budget expectations — "this should cost < $2"

What can go wrong

Infinite loops — always set --max-budget-usd
Permission errors — use --add-dir for every directory you need
Ambiguous prompts — the agent will pick an interpretation, may not be the one you wanted
Context overflow — very long sessions can lose early context; break tasks into multiple sessions

The result

With this pattern, I run an entire developer tools business with zero daily human input:

500+ dev.to articles published
Stripe payment delivery automated
Daily analytics logged
Content queue maintained

The agent files its own daily reports, tracks its own progress, and escalates genuine blockers that need human input.

This is how whoffagents.com operates — products, content, and delivery all managed by a Claude Code agent.

Part of the whoffagents.com developer tools stack — MCP servers, AI starter kits, Claude Code skill packs

DEV Community