Consult Agents

Get validation, analysis, or research feedback from Gemini and Codex before Claude Code implements.

Core principle: Claude Code implements. External agents validate and research.

When to Use

•Validate an approach before implementing
•Research a topic from multiple perspectives
•Get code review feedback on a design
•Confirm understanding of requirements
•Aggregate opinions on trade-offs

Model Requirements

CRITICAL: Always use top-tier models. Never use smaller/cheaper models.

CLI	Model Flag	Required Model
Codex	`-m gpt-5.2`	gpt-5.2 (never o4-mini, gpt-4o, etc.)
Gemini	`-m gemini-3-pro`	gemini-3-pro (never smaller models)

Quick Reference

Task Type	Gemini Mode	Codex Mode
Analyze code/files only	`gemini -m gemini-3-pro "..."`	`-m gpt-5.2 --sandbox read-only`
Research with web access	`gemini -m gemini-3-pro "..."` (has web search)	`-m gpt-5.2 --sandbox danger-full-access`
Deep repo analysis	tmux interactive	`-m gpt-5.2 --sandbox read-only`

Workflow

code

1. Prepare question/context (what you need validated)
2. Query Gemini → capture response
3. Query Codex → capture response
4. Synthesize: agreements, disagreements, insights
5. Claude Code decides and implements

Parallel Execution

Run both agents simultaneously for faster feedback:

bash

# Terminal 1: Gemini
gemini -m gemini-3-pro "Analyze this approach: [context]. What are the trade-offs?" 2>/dev/null

# Terminal 2: Codex
echo "Analyze this approach: [context]. What are the trade-offs?" | \
  codex exec -m gpt-5.2 --sandbox read-only --skip-git-repo-check 2>/dev/null

Or use background jobs:

bash

gemini -m gemini-3-pro "..." > /tmp/gemini-response.txt 2>/dev/null &
echo "..." | codex exec -m gpt-5.2 --sandbox read-only --skip-git-repo-check > /tmp/codex-response.txt 2>/dev/null &
wait

Synthesis Template

After gathering responses, synthesize:

markdown

## Agent Responses

### Gemini
[Summary of Gemini's perspective]

### Codex
[Summary of Codex's perspective]

## Synthesis

**Agreements:**
- [What both agents agree on]

**Disagreements:**
- [Where they differ and why]

**Key Insights:**
- [Unique valuable points from either]

## Decision
[Claude Code's decision based on aggregated feedback]

Prompt Templates

Validation

code

Review this [design/approach/code]:

[CONTEXT]

Questions:
1. Is this approach sound?
2. What are the risks?
3. What would you do differently?

Research

code

Research [TOPIC] focusing on:
1. Current best practices
2. Common pitfalls
3. Recommended tools/libraries

Context: [Why you need this]

Code Review

code

Review this code for:
- Correctness
- Performance concerns
- Security issues
- Maintainability

[CODE]

Common Mistakes

Mistake	Fix
Asking agents to implement	They validate; Claude Code implements
Not providing enough context	Include relevant code, constraints, goals
Ignoring disagreements	Disagreements reveal important trade-offs
Blindly following consensus	Use judgment; agents can both be wrong

When NOT to Use

•Simple, obvious implementations
•Time-critical changes (adds latency)
•Trivial questions (overkill)