Building and Configuring AI Agents: Best Practices and Implementation Guide

AI agents can automate research, qualification, and audience building workflows. This guide covers best practices for configuring agents, integrating them with plays and audiences, and optimizing their performance.

Core Workflow Overview

Follow this comprehensive workflow for research-based snippets and dynamic audience building:

1. Create Company or Person-level Agents with targeted research questions and specific guidance

2. Test and iterate Agents until they reliably return values before implementing in workflows

3. Use Plays with "AI qualification" actions to automatically execute Agents and qualify companies

4. Create Smart Snippets that reference Agent outputs via template variables with clear fallback text

5. Build dynamic Audiences filtered by Agent-generated tags/fields

6. Re-run Plays on seed Audiences at desired intervals for periodic monitoring

7. Use "Generate examples" feature to refine snippet outputs

Agent Configuration Best Practices

Question Types and Structure

Design your agent questions based on your research objectives:

• True/False questions for qualification and detection (e.g., "Does this company have a recent product launch that mentions AI?")

• Open text questions for detailed information capture (e.g., "What is the most recent product launch name?", "Link to the announcement source", "Launch date")

• Combined approach: Use True/False for detection plus open text for details (URLs, snippets, dates) within the same agent

Guidance Configuration

Include clear guidance to improve agent reliability:

• Fallback logic: "If uncertain, answer False"

• Keyword detection: "Treat 'AI' as present if post includes terms like 'AI', 'artificial intelligence', 'LLM', 'GPT', or 'machine learning'"

• Data capture instructions: "Capture canonical URL, copy short quote showing mention, record post date. If multiple exist, use most recent"

Agent Architecture Patterns

Choose the right architecture based on your use case:

• Single Agent with Multiple Questions: Multiple questions within one agent step share the same context window and are processed by the same model instance

• Sequential Agent Routing: Create separate agent steps for different signals and route companies through them in order of importance to optimize credit usage

• Hybrid Approach: Combine multiple questions in one agent step plus routing to multiple agents for different scenarios

Play Integration for Automated Qualification

Configure Plays to automatically execute agents and process results:

1. Add "AI qualification" action in Plays with your research Agent

2. Set qualification rules based on Agent responses (e.g., proceed when True)

3. Configure downstream actions to tag qualified companies or update custom fields

4. Optionally sync results to your CRM

Dynamic Audience Creation

Build audiences that automatically update based on agent results:

• Create audiences filtered by Agent-generated tags/fields (e.g., recent_launch = true)

• These audiences automatically update as new companies are qualified

• This enables automated audience building based on real-time external data rather than static uploads

Smart Snippet Integration

Reference agent outputs in your messaging with proper fallbacks:

• Use template variables to reference Agent outputs with clear fallback text

• Store multiple data points from agents (URLs, snippets, dates) in company fields

• Example prompt: "Research indicates {{agent_recent_launch_name}} was recently announced by {{company_name}}. Write a 1–2 sentence hook congratulating them... If no launch is found, write a neutral industry-relevant hook."

• Use the "Generate examples" feature to refine snippet outputs

Example Implementation: Product Launch Detection

Here's a practical example for detecting AI-related product launches:

Agent Setup

• Primary Question (True/False): "Does this company have a recent product launch that mentions AI?"

• Additional capture fields: Launch URL, AI mention snippet, launch date

• Agent Guidance: Search for company launches, look for AI-related terms (AI, artificial intelligence, LLM, GPT, machine learning)

Integration Workflow

1. Use AI qualification action in Plays

2. Tag qualified companies (e.g., ai_launch = true)

3. Create dynamic Audiences filtering on this tag

4. Reference Agent variables in Smart Snippets with fallbacks

Credit Optimization

Manage credit usage effectively:

• Each Agent question per company uses 1 credit

• Plan credit usage when running Agents at scale

• Optimization tip: When looking for multiple signals, use sequential agent routing in order of importance to save credits by stopping at the first match

Periodic Monitoring

Keep your data current with regular updates:

• Re-run Plays on seed Audiences at desired intervals for periodic monitoring

• Note: Scheduled agent runs aren't natively supported, so manual re-runs are the current workaround