How to manage Claude Code, Codex, OpenClaw & Hermes from iPhone (2026 playbook)

Deploying an AI agent is the easy part. The real skill is managing agents running across multiple servers while you're away from your desk. Here's the practical playbook.

Session Persistence with tmux

The #1 rule: never run a long-lived agent in a bare SSH session. Use tmux:

tmux new -s agent-task

Start your agent inside tmux. If your SSH connection drops — phone goes to sleep, train enters a tunnel — the agent keeps running. Reconnect later and reattach:

tmux attach -t agent-task

This works from any device. Start the agent from your laptop, check progress from your phone over SSH later.

Monitoring Progress

AI agents produce a lot of output. Here's how to stay on top of it:

• Scroll back — your iOS terminal app should support full scroll history. Swipe up to review what the agent did while you were away.
• Git diff — after the agent finishes, run git diff to see exactly what changed. This is the fastest way to audit the work.
• Test suites — if your project has tests, run them. A passing suite is the best signal that the agent did its job.

Multi-Agent Workflows

Running agents across multiple servers? Keep it organized:

1. One tmux session per task — name them descriptively: tmux new -s fix-auth, tmux new -s migrate-db
2. One project per agent — don't let two agents edit the same codebase simultaneously
3. Use branches — each agent works on its own git branch. Review and merge when satisfied.

Error Handling

Agents can get stuck. Common scenarios and fixes:

• Agent loops on a failing test — intervene via the terminal. Type a clarification or redirect its approach.
• Agent runs out of context — restart with a more focused task description.
• Permission errors — the agent inherits your SSH user's permissions. Check file ownership and sudo access.

Security Considerations

An AI agent has the same access as your SSH session. This means:

• It can read all files your user can read
• It can execute any command your user can run
• It can access any API key in the environment

Use least-privilege principles. Run agents as a dedicated user with limited permissions when possible.

Skill triage for self-improving agents

The new wrinkle in 2026 is self-improving agents like Hermes that auto-generate skill files as they work. A live Hermes runtime commonly produces five to ten new skills per active day, and the agent's self-evaluation is documented to be over-confident — without human triage, the skill library grows noisy fast. The 30-second daily routine: open the skills view, promote the obvious wins, demote or delete the noisy ones, lock the ones you have hand-tuned. See manage Hermes skills from iPhone in one touch.

The iPhone Advantage

Your phone is the ideal management interface. You don't need the full power of a laptop to check a git diff, approve a PR, restart a stuck agent, or triage a skill — terminal access to all your servers plus a one-touch skill view is enough. Check on agents between meetings, on the bus, or from bed.

New to this? Start with what an AI agent on iPhone actually is, then compare iOS clients in Termius vs Blink vs Onepilot. For framework agents specifically, Onepilot packages the deploy wizard and channel routing — see OpenClaw on iPhone, Hermes on iPhone, and Paperclip multi-agent orchestrator.