Orchestrate Agent Clis

Orchestrate Agent CLIs

Execute different AI agent CLIs to get alternative perspectives, compare approaches, or leverage specific model strengths.

Never truncate agent output

• Run every command in this skill raw. Never pipe it through anything that drops lines
• Forbidden: tail, head, sed -n, grep/rg filters, cut, wc -l, | less, --quiet-style flags
• Truncated output is gone. The agent's answer is the deliverable, and the cut part is unrecoverable without paying for a full rerun
• Output too long for chat: redirect the FULL output to a file in the scratchpad dir, then read the file with the read tool
  • codex exec --skip-git-repo-check "<prompt>" < /dev/null >/path/to/scratchpad/codex.log 2>&1
• Same rule for model discovery (agy models, cursor-agent models, pi --list-models)

Always redirect stdin

• Append < /dev/null to EVERY invocation in this skill, unless you are deliberately piping input
• Cause: these CLIs read the prompt from stdin when it is piped. An open pipe that never closes means they block waiting for EOF
• codex exec is the worst case: it prints Reading additional input from stdin..., never processes the argument prompt, and produces zero output. Signature: process alive, ~0% CPU, no output
• agy -p, cursor-agent -p, claude -p answer the prompt but then fail to exit. Work is done, process hangs
• < /dev/null fixes all four (verified: exit 0)

# hangs
codex exec --skip-git-repo-check "<prompt>"

# correct
codex exec --skip-git-repo-check "<prompt>" < /dev/null

No memory between runs

• Every invocation starts a fresh instance with a clean context. No memory of any previous run, yours or another agent's
• Write each prompt standalone. Include the file paths, task, constraints, and any earlier findings the agent needs to do the work
• Forbidden in prompts: "as we discussed", "continue with", "the previous answer", "your earlier review", "remember that", or any reference to a prior run
• Content from a previous run that the agent needs: paste it inline as data, not as shared history. Never assume it can look it up

Defaults

• If user names an agent but not provider, model, or effort, use that CLI's default
• Do not ask which provider, model, or effort to use unless user explicitly asks to choose
• Do not pause to recommend alternatives when the requested agent can run with defaults
• Ask only for missing task inputs that block execution, such as absent file paths or destructive write intent

Available CLIs

agy (Antigravity CLI)

Google's Antigravity CLI, the successor to Gemini CLI.

agy -p "<prompt>" < /dev/null

Has internet access. Good for research and current information.

--effort low|medium|high controls reasoning depth.

Quota: two separate buckets. Flash refills ~every 5h; Pro and oss refill weekly. Prefer Flash; save Pro/oss for a drained Flash bucket or a genuinely different model.

Discover models (use when a --model name fails or may be stale):

agy models

Codex

codex exec --skip-git-repo-check "<prompt>" < /dev/null

No model-list subcommand. Set with -m <model> (e.g. -m gpt-5.6-sol). Discover current OpenAI model IDs at https://platform.openai.com/docs/models or the interactive /model picker in codex (no subcommand).

Claude Code

claude -p "<prompt>" < /dev/null

Full-featured Claude agent with tool access.

No model-list subcommand. --model takes an alias (opus, sonnet, haiku) or a full ID (e.g. claude-opus-4-8). Discover current IDs at https://docs.claude.com/en/docs/about-claude/models/overview.

Cursor Agent

cursor-agent --model=MODEL -p "<prompt>" < /dev/null

Available models (not extensive, favorites):

Note: If these models fail, get the list of available models with this command (no truncation, see "Never truncate agent output"):

cursor-agent models

Pi

pi -p "<prompt>" < /dev/null

Multi-provider agent (Anthropic, OpenAI, Google, xAI, etc)

Model and thinking mode are optional for Pi. Omit --model and --thinking unless user requested them.

• --thinking <off|minimal|low|medium|high|xhigh> controls reasoning depth.
• --tools read,grep,find,ls for read-only runs.
• --continue / --resume to reuse sessions.
• --list-models [search] to discover models.

File References

Reference files using @ prefix with relative paths:

codex exec --skip-git-repo-check "Review @src/components/Button.tsx" < /dev/null
agy -p "Create tests for @src/utils/helpers.ts" < /dev/null
cursor-agent --model=composer-2.5 -p "Refactor @src/api/client.ts" < /dev/null
pi -p "Review @src/api/client.ts" < /dev/null

Task Guidelines

Read-Only Tasks

For questions, plans, reviews, or analysis, instruct agents not to make file changes:

cursor-agent --model=gemini-3.1-pro -p "Analyze the architecture in @src/core/ - explain only, do not modify files" < /dev/null

Comparison Workflows

Get multiple perspectives on the same problem. Send the identical standalone prompt to each CLI. Neither run sees the other, so you do the comparing:

agy -p "How would you implement feature X?" < /dev/null
codex exec --skip-git-repo-check "How would you implement feature X?" < /dev/null