November 18, 2026 · TermBridge · 4 min read

Watch Self-Improving Agent Harness Runs from iPhone

Watch Self-Improving Agent Harness Runs from iPhone

Self-improving agents run in loops. They attempt a task, score the result, adjust and try again. One loop can take an hour, and the run can stall on a permission prompt or a failed tool call while you are away from the desk. You only find out when you sit back down.

We (GREP.AI, YC F26) are open-sourcing AgentRun and turning it into a full managed agents platform. […] It includes what we originally built for our own agent harness: Self-improving agents, Web-grounded workflows, A range of sandboxes and computer environments, Thousands of tools and integrations, Detailed analytics on how your agents behave. You can bring your own agent. AgentRun works with Pi, Claude Agents, Codex, OpenCode, and any other harness that supports ACP.

What TermBridge does here: if you self-host AgentRun or run Claude Code, Codex or OpenCode directly on your Mac, the work happens in a terminal. TermBridge mirrors that exact shell, whether it’s iTerm2, Terminal.app, a VS Code, Zed or Cursor integrated terminal, or a tmux pane, to your iPhone or iPad. You can watch output scroll and type replies. Pairing is a QR scan approved on the Mac, and the stream is end-to-end encrypted with X25519 and AES-256-GCM, so the relay only sees ciphertext.

A laptop on a desk showing colorful source code in a dark editor

Catch the loop when the self-improvement step stalls

The pitch in the AgentRun Show HN thread is agents that get better across runs. In practice that means many iterations of the same harness, and each iteration can hit an approval prompt or a tool error. Analytics will tell you afterward what went wrong. A live terminal on your phone tells you while it is happening, so you can approve the command or press Ctrl-C before a bad iteration burns another round of OpenRouter tokens.

Bring your own harness, keep your own terminal

AgentRun says it works with Pi, Claude Agents, Codex and OpenCode over ACP. Plenty of people will still test a harness locally before wiring it into a platform. If Codex runs in one tmux window and Claude Code in another, TermBridge shows both as separate terminals on the phone, and you switch between them with a tap. The setup is similar to the one in watching a mixed Claude Code, Codex and Aider setup, just with a self-improving loop on top.

iPhone screen listing several active Mac terminal sessions running different agents

Self-hosted on a spare Mac, checked from the couch

The OP offers a hosted console or a self-hosted install with an OpenRouter key. If you choose self-hosting on a Mac mini in the corner, you have no screen nearby. Pair your iPhone and iPad to that Mac, and pair other Macs too if you split sandboxes across machines. Then switch between them from the device list. Because traffic is encrypted end to end, the agent’s logs, keys printed by mistake and tool output stay unreadable to the relay.

Discussed on Hacker News · 138 pts · 41 💬

Discussed on

← Back to Blog