FrancoStino/delegating-to-agents

Delegating to Agents

When to Use

• Use when work should be handed to another AI agent with a complete prompt and progress checks.
• Use when you need to relay instructions to terminal or TUI agents without losing context.

Which agent to pick

• Coding (default) → Codex CLI. Strongest coding agent, especially for complex, long-running SWE tasks. It's on an unlimited-usage plan — effectively unlimited, don't ration it.
• Most other tasks → Pi Agent (pi in a cmux terminal). All Pi agents run opus-4.8-fast via OpenRouter at xhigh reasoning effort.
• Frontend / design → Pi. Opus 4.8 Fast beats Codex on UI, styling, design.
• Heavy multi-step work: you as orchestrator + Codex CLI executing in a right-hand cmux pane is a solid default setup.

Sending prompts to a TUI agent

1. ONE single line — never newlines in the message body. In a TUI, newline = Enter: a multi-line prompt submits at the first line and the rest arrives as fragmented mid-turn steering messages. Use ". " or "; " instead of line breaks, then one explicit enter. For long instructions, write them to a file and send: read /tmp/task.md and follow it.
2. Wrap the prompt in plain double quotes — NEVER escaped. cmux send --surface surface:N "your prompt". The recurring bug is emitting \" — in bash that's literal-broken and dies with unexpected EOF. Inside the prompt, avoid apostrophes and literal double quotes (write "dont", "wont", "lets"); rephrase instead of escaping. If a send failed, the cause was the escaped \", not the quote type.
3. Exact command names: cmux send --surface surface:N then cmux send-key --surface surface:N enter. There is NO send-surface or send-key-surface.

Polling

Keep sleeps SHORT: start at 3-5s, re-check, repeat. Don't sleep 30. Pi and Hermes (opus-4.8-fast) launch and respond within seconds; scale up only for genuinely heavy tasks. After every check, send the user a one-line status: what the agent is doing and whether it's on track.

Claude Code note: after it finishes, it may prefill a predicted next user message — that draft is Claude, not the user.

Remote VPS

SSH in first and launch the agent ON the VPS (e.g. codex --yolo), then drive that on-box agent. Don't run an agent locally and have it SSH for every step.

The 4 agents (background reference)

All four use the portable SKILL.md standard; project skills win over global.

• Pi (pi.dev, open-source TS): minimal read/write/edit/bash core, self-extends via TS extensions; true BYOK; best-in-class session branch/fork/resume. Skills: ~/.pi/agent/skills/.
• Codex CLI (OpenAI, Rust): fastest startup; kernel-level sandboxing; codex exec for CI; reads AGENTS.md. Skills: ~/.codex/skills/.
• Claude Code (Anthropic, TS): deepest Claude integration, .claude/ conventions, live skill hot-reload. Skills: ~/.claude/skills/.
• Hermes (Nous Research, Python): persistent autonomous agent — cross-session memory, built-in scheduler, 40+ tools; can orchestrate the other CLIs as workers. Skills: ~/.hermes/skills/.

Driving interactive CLIs

• Codex, Pi, OpenCode: need pty=true.
• Claude Code: prefer claude --print --permission-mode bypassPermissions (no PTY).

Limitations

• Adapted from davidondrej/skills; verify local paths, tools, credentials, and agent features before acting.
• For commands, remote access, scheduling, browser automation, or file-changing workflows, get explicit user approval and confirm the target environment first.