Agent Selection Guide

This guide helps you choose the right AI agents for your Aragora stress-tests based on task type, cost, and capability requirements.

Perspective coverage note: Mistral adds an EU lens, and Chinese models like DeepSeek, Qwen, and Kimi provide a Chinese perspective (use the providers and keys listed below).

Available Agents

Primary Providers (Direct API)

Agent ID	Provider	Model	Best For	Cost
`anthropic-api`	Anthropic	claude-opus-4-5-20251101	Code review, reasoning	$$
`openai-api`	OpenAI	gpt-5.2	General tasks, creativity	$$
`gemini`	Google	gemini-3-pro-preview	Long context, analysis	$
`mistral-api`	Mistral	mistral-large-2512	European compliance, multilingual	$$
`grok`	xAI	grok-4-latest	Real-time knowledge	$$

OpenRouter Providers (Fallback/Alternative)

Agent ID	Model	Best For	Cost
`openrouter`	model parameter (default: deepseek/deepseek-chat-v3-0324)	Fallback when primary fails	Varies
`deepseek`	deepseek/deepseek-reasoner	Code, math, reasoning	$
`deepseek-r1`	deepseek/deepseek-r1	Chain-of-thought reasoning	$
`mistral`	mistralai/mistral-large-2411	Fast, high-quality reasoning	$$
`qwen`	qwen/qwen3-max	Multilingual, code	$
`qwen-max`	qwen/qwen3-max	Flagship reasoning	$$
`llama`	meta-llama/llama-3.3-70b-instruct	General, open weights	$
`yi`	01-ai/yi-large	Chinese/English	$

Cost Legend: $ = Low ($0.001-0.01/1K tokens), $$ = Medium ($0.01-0.05/1K), $$$ = High ($0.05+/1K)

Local Providers (No API Key)

Agent ID	Model	Best For	Cost
`ollama`	Local Ollama model	Air-gapped/private deployments	$
`lm-studio`	Local LM Studio model	Desktop/local inference	$

Environment variables:

export OLLAMA_HOST=http://localhost:11434
export OLLAMA_MODEL=llama2
export LM_STUDIO_HOST=http://localhost:1234

CLI usage:

aragora ask "Review this policy" --agents ollama
aragora ask "Summarize this spec" --agents lm-studio

Python autodetection:

from aragora.agents import LocalLLMDetector

status = await LocalLLMDetector().detect_all()
if status.any_available:
    print(status.recommended_server, status.recommended_model)

API endpoints:

curl -s http://localhost:8080/api/agents/local
curl -s http://localhost:8080/api/agents/local/status

Task-Based Recommendations

Code Review

Recommended: anthropic-api,openai-api

git diff main | aragora review --agents anthropic-api,openai-api

Why:

Anthropic excels at code understanding and security analysis
OpenAI provides creative edge case detection
Consensus between them = high confidence findings

Budget alternative: anthropic-api,deepseek

DeepSeek V3 is excellent at code for 1/10th the cost

Architecture Design

Recommended: anthropic-api,openai-api,gemini

aragora ask "Review this microservices architecture" \
  --agents anthropic-api,openai-api,gemini \
  --rounds 3

Why:

Three diverse perspectives catch more issues
Gemini handles long architecture documents well
Multiple rounds allow deeper exploration

Compliance Audits

Recommended: anthropic-api,mistral-api

aragora gauntlet policy.md --agents anthropic-api,mistral-api --persona gdpr

Why:

Mistral is trained with European data/compliance focus
Anthropic provides strong reasoning for legal interpretation
Both have strong safety training

Quick Validation

Recommended: anthropic-api (single agent)

aragora review --demo  # Or single agent for speed
aragora review --agents anthropic-api

Why:

Fastest response time
Anthropic alone catches most critical issues
Use for early-stage development feedback

High-Stakes Decisions

Recommended: anthropic-api,openai-api,gemini,mistral-api

aragora gauntlet critical_spec.md \
  --agents anthropic-api,openai-api,gemini,mistral-api \
  --profile thorough

Why:

Four perspectives maximize coverage
Different training data catches different blind spots
Worth the cost for critical decisions

Cost Optimization

Single API Key Strategy

If you only have one API key:

# Anthropic only
export ANTHROPIC_API_KEY=your_key
export ARAGORA_DEFAULT_AGENTS=anthropic-api
aragora review

# OpenAI only
export OPENAI_API_KEY=your_key
export ARAGORA_DEFAULT_AGENTS=openai-api
aragora review

Budget-Conscious Setup

Use OpenRouter for cost-effective multi-agent:

export OPENROUTER_API_KEY=your_key
aragora review --agents deepseek,qwen

Estimated costs per 10K token stress-test:

anthropic-api,openai-api: ~$0.30
deepseek,qwen: ~$0.03 (10x cheaper)

Automatic Fallback

Aragora auto-falls back to OpenRouter on rate limits:

# Primary keys + fallback
export ANTHROPIC_API_KEY=your_key
export OPENAI_API_KEY=your_key
export OPENROUTER_API_KEY=fallback_key

# If primary hits rate limit, falls back automatically
aragora review --agents anthropic-api,openai-api

If a provider key is missing and OPENROUTER_API_KEY is set, Aragora will substitute OpenRouter models to keep the debate running:

anthropic-api -> anthropic/claude-3.5-sonnet
openai-api -> openai/gpt-4o-mini
gemini -> google/gemini-2.0-flash-exp:free
grok -> x-ai/grok-2-1212
mistral-api -> mistralai/mistral-large-2411

Capability Matrix

Capability	Anthropic	OpenAI	Gemini	Mistral	DeepSeek
Code Understanding	★★★★★	★★★★☆	★★★★☆	★★★★☆	★★★★★
Security Analysis	★★★★★	★★★★☆	★★★☆☆	★★★★☆	★★★☆☆
Reasoning Depth	★★★★★	★★★★★	★★★★☆	★★★★☆	★★★★☆
Long Context	★★★★☆	★★★★☆	★★★★★	★★★☆☆	★★★★☆
Multilingual	★★★☆☆	★★★★☆	★★★★☆	★★★★★	★★★★☆
Creativity	★★★★☆	★★★★★	★★★☆☆	★★★☆☆	★★★☆☆
Safety/Refusals	★★★★★	★★★★☆	★★★★☆	★★★★☆	★★★☆☆
Response Speed	★★★★☆	★★★☆☆	★★★★★	★★★★☆	★★★★☆

Role Assignment

Aragora auto-assigns roles based on agent order:

aragora ask "Design auth system" --agents anthropic-api,openai-api,gemini

Position	Role	Best Agent Type
1st	Proposer	Strong reasoning (Anthropic, OpenAI)
2nd-n-1	Critic	Detail-oriented (Anthropic, Mistral)
Last	Synthesizer	Balanced (OpenAI, Gemini)

Agent Specification Formats

Recommended - Pipe format (provider|model|persona|role):

# Full specification with all fields
--agents "anthropic-api|claude-opus|philosopher|proposer,openai-api|gpt-4o|skeptic|critic"

# Provider and role only (most common)
--agents "anthropic-api|||proposer,openai-api|||critic,gemini|||synthesizer"

# Provider with persona (role auto-assigned)
--agents "anthropic-api||philosopher|,openai-api||skeptic|"

Legacy - Colon format (backward compatible):

# Role assignment
--agents anthropic-api:proposer,openai-api:critic,gemini:synthesizer

# Persona assignment (non-role words treated as personas)
--agents anthropic-api:philosopher,openai-api:skeptic

Format	Example	Provider	Model	Persona	Role
Pipe (full)	`anthropic-api\|claude-opus\|phil\|critic`	anthropic-api	claude-opus	phil	critic
Pipe (role)	`anthropic-api\|\|\|critic`	anthropic-api	default	-	critic
Pipe (persona)	`anthropic-api\|\|philosopher\|`	anthropic-api	default	philosopher	auto
Legacy (role)	`anthropic-api:critic`	anthropic-api	default	-	critic
Legacy (persona)	`anthropic-api:philosopher`	anthropic-api	default	philosopher	auto
Plain	`anthropic-api`	anthropic-api	default	-	auto

Valid roles: proposer, critic, synthesizer, judge

Environment Variables

# Primary providers
export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...
export GEMINI_API_KEY=AI...
export MISTRAL_API_KEY=...
export XAI_API_KEY=...

# Fallback provider
export OPENROUTER_API_KEY=sk-or-...

Recommendations by Use Case

Use Case	Agents	Rounds	Profile
Quick PR review	`anthropic-api,openai-api`	2	-
Security audit	`anthropic-api,openai-api,mistral-api`	3	`thorough`
Architecture review	`anthropic-api,openai-api,gemini`	3	`thorough`
GDPR compliance	`anthropic-api,mistral-api`	2	`policy`
Code refactoring	`anthropic-api,deepseek`	2	`code`
Budget review	`deepseek,qwen`	2	`quick`
High-stakes decision	All 4 primary	4	`thorough`

Troubleshooting

"No API keys configured"

Set at least one provider key:

export ANTHROPIC_API_KEY=your_key

Rate limiting

Add OpenRouter fallback:

export OPENROUTER_API_KEY=fallback_key

Slow responses

Use fewer agents
Use --rounds 1 for faster results
Use gemini (fastest response time)

Inconsistent results

Add more agents for consensus
Use --rounds 3 or more
Prefer Anthropic for consistent reasoning

Environment Variables - Full API key reference
Gauntlet Mode - Stress testing configuration
API Reference - Programmatic agent selection

Available Agents​

Primary Providers (Direct API)​

OpenRouter Providers (Fallback/Alternative)​

Local Providers (No API Key)​

Task-Based Recommendations​

Code Review​

Architecture Design​

Compliance Audits​

Quick Validation​

High-Stakes Decisions​

Cost Optimization​

Single API Key Strategy​

Budget-Conscious Setup​

Automatic Fallback​

Capability Matrix​

Role Assignment​

Agent Specification Formats​

Environment Variables​

Recommendations by Use Case​

Troubleshooting​

"No API keys configured"​

Rate limiting​

Slow responses​

Inconsistent results​

Related Documentation​

Available Agents

Primary Providers (Direct API)

OpenRouter Providers (Fallback/Alternative)

Local Providers (No API Key)

Task-Based Recommendations

Code Review

Architecture Design

Compliance Audits

Quick Validation

High-Stakes Decisions

Cost Optimization

Single API Key Strategy

Budget-Conscious Setup

Automatic Fallback

Capability Matrix

Role Assignment

Agent Specification Formats

Environment Variables

Recommendations by Use Case

Troubleshooting

"No API keys configured"

Rate limiting

Slow responses

Inconsistent results

Related Documentation