Delegate complex tasks to specialized AI coding agents

Delegates work to another AI agent - handing it a complete prompt, sending it correctly to a terminal or TUI agent without losing context, and checking on its progress. Covers agent selection, how to send prompts to a TUI agent without breaking the terminal, polling cadence, and driving agents on a remote VPS.

Use it when work should be handed to another AI agent with a complete prompt and progress checks, or when instructions need to be relayed to terminal or TUI agents without losing context. For coding, default to Codex CLI - the strongest coding agent, especially for complex, long-running SWE tasks, on an unlimited-usage plan that doesn't need rationing. For most other tasks, use Pi Agent (pi in a cmux terminal; all Pi agents run opus-4.8-fast via OpenRouter at xhigh reasoning effort), and specifically prefer Pi for frontend or design work, where Opus 4.8 Fast beats Codex on UI, styling, and design. For heavy multi-step work, a solid default is acting as orchestrator yourself while Codex CLI executes in a right-hand cmux pane.

Sending a prompt to a TUI agent has three hard rules: send it as ONE single line, never newlines in the message body - in a TUI a newline is Enter, so 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, or for long instructions write them to a file and send a pointer like "read /tmp/task.md and follow it". Wrap the prompt in plain double quotes, never escaped - cmux send --surface surface:N "your prompt" - since an escaped \" breaks bash with unexpected EOF; avoid apostrophes and literal double quotes inside the prompt (write "dont", "wont", "lets") rather than escaping them. Use the exact commands cmux send --surface surface:N followed by cmux send-key --surface surface:N enter - there is no send-surface or send-key-surface. When polling for a response, keep sleeps short (start at 3-5s and re-check, never sleep 30, since Pi and Hermes launch and respond within seconds and only genuinely heavy tasks need longer), and send the user a one-line status after every check. On a remote VPS, SSH in first and launch the agent on the VPS itself (e.g. codex --yolo), then drive that on-box agent, rather than running an agent locally and having it SSH for every step.

Covers four interoperable CLI agents, all using the portable SKILL.md standard where project skills win over global ones: Pi (pi.dev, open-source TypeScript) with a minimal read/write/edit/bash core that self-extends via TS extensions, true BYOK, and best-in-class session branch/fork/resume, skills in ~/.pi/agent/skills/; Codex CLI (OpenAI, Rust) with the fastest startup, kernel-level sandboxing, codex exec for CI, and AGENTS.md support, skills in ~/.codex/skills/; Claude Code (Anthropic, TypeScript) with the deepest Claude integration, .claude/ conventions, and live skill hot-reload, skills in ~/.claude/skills/; and Hermes (Nous Research, Python), a persistent autonomous agent with cross-session memory, a built-in scheduler, 40+ tools, and the ability to orchestrate the other CLIs as workers, skills in ~/.hermes/skills/. Codex, Pi, and OpenCode need pty=true to drive interactively; Claude Code is instead best driven with claude --print --permission-mode bypassPermissions (no PTY). Claude Code may also prefill a predicted next user message after finishing - that draft is Claude's own prediction, not the user's.

Orchestrating agents or developers who need to reliably hand off coding, frontend, or heavy multi-step work to the right specialized CLI agent - locally or on a remote VPS - without breaking prompt delivery or losing track of progress.