Skip to main content

Consult Tool

Consult is a built-in AdaL tool that lets your AI agent ask one or more AI models for a second opinion while staying on its primary model. It compares different perspectives, leverages specialized model expertise, and continues the workflow automatically.

Quick Start​

Ask AdaL in plain language — the agent decides autonomously when to consult based on its system prompt guidance. You don't have to script it.

Let the Agent Decide​

Review this implementation and use Consult to get a second opinion
from a different model before making changes.

Consult One Model​

Name a specific model for targeted specialist feedback:

Review this function for correctness, edge cases, and maintainability.
Consult claude-sonnet-4-6 for a senior code reviewer's perspective.

For one model, Consult returns that model's response directly — no header, no separator.

Consult Multiple Models​

Ask for multiple perspectives when diverse opinions matter:

Should this service use PostgreSQL or DynamoDB?
Consult both claude-sonnet-4-6 and gpt-5.4 for independent assessments.

For multiple models, Consult labels each response with its model tag:

## anthropic-claude-sonnet-4-6
PostgreSQL is better for complex queries...

---

## openai-gpt-5.4
DynamoDB for scale, PostgreSQL for query flexibility...

The agent reads the perspectives, notes where they agree and disagree, and makes a more informed decision.

Under the Hood — Minimal Prompt, Efficient Token Usage​

Consult uses a simple pass-through prompt design. The system prompt is just a plain string — no complex multi-message structure, no extra overhead. This keeps every call lightweight and token-efficient:

# Backend tool call — what AdaL sends to each model
consult(
    prompt="Should this service use PostgreSQL or DynamoDB?",
    model_tags=["anthropic-claude-sonnet-4-6", "openai-gpt-5.4"],
    system_instruction="Review independently and state your assumptions.",
)

# What each model actually receives (single user-role string):
# [System]: Review independently and state your assumptions.
#
# Should this service use PostgreSQL or DynamoDB?

The system_instruction is prefixed as [System]: ... and combined with the prompt into a single string. No multi-message API calls, no extra formatting — just one clean request per model. This means:

  • Minimal tokens spent on structure — almost every token goes to the actual question
  • Same simple format for all providers — no API-specific message handling
  • Fast and cheap — a focused 50-token prompt × 2 models costs less than a single long conversation turn