Creating an AI Agent — Step 1: Build

The Build step is where you define what your agent is, what data it processes, and which LLM powers it. Get these decisions right and the rest of the wizard follows naturally.

Written By Kevin Lawrie

Last updated About 3 hours ago

Getting started

Go to AI Agents in the left sidebar and click Create AI Agent. The creation wizard opens with three steps: Build, Prompt, and Actions.

Agent Name and Description

Give your agent a clear, descriptive name — you'll assign multiple agents to Signals over time and need to tell them apart at a glance.

Good names:

"ICP Verifier — Contacts"
"Buyer Intent — LinkedIn Comments"
"Cold Email Opener — Contacts"
"YouTube Comment Sentiment"

The description is optional but useful for team accounts where others may be working with your agents.

Data Target

This is the most important decision in the Build step. The data target determines:

What your agent reads (posts, comments, contact profiles, videos)
Which variables are available in your prompt
Which actions are available in your conditional rules

Data target	Best for
Contacts	Intent, Pain identification, ICP scoring, persona matching, personalised copy generation — runs against enriched contact records
Posts	Qualifying LinkedIn posts by content and intent, extracting post authors as contacts when relevant
LinkedIn Comments	Qualifying individual commenters by what they said — saving only relevant commenters as enriched contacts
YouTube Videos	Summarising video content, extracting themes, writing analysis to video records
YouTube Comments	Finding signal in YouTube comment sections — intent, sentiment, competitive mentions

⚠️ Note: You cannot change the data target after creating an agent. If you need a different target, create a new agent.

LLM Provider and Model

AI Agents use your own LLM provider account — bring your own key (BYOK). Only providers you have connected in Integrations appear here.

Supported providers: OpenAI, Anthropic, Gemini, xAI, Perplexity

To connect a provider, go to Integrations → filter by AI & ML → click Configure on your chosen provider → enter your API key → Save.

Choosing a model

You have full control over which model powers each agent. General guidance:

High-volume classification (TRUE/FALSE, category labels on posts or comments) — use a faster, cheaper model. You may be processing thousands of items per Signal run.
Nuanced analysis or generation (ICP scoring with reasoning, personalised copy) — use a more capable model where output quality matters more than cost per call.

LLM costs are charged directly by your provider and do not come from your platform credits. Platform cost is a flat -0.1 credits per agent job regardless of model.

Advanced LLM Settings

Click Advanced LLM Settings to expand optional configuration. These are for users who want fine-grained control over model behaviour. Defaults work well for most agents.

		Setting What it does When to adjust
Temperature (0–2)	Controls creativity vs consistency. Lower = more predictable outputs.	Lower it (0.1–0.3) for classification agents where you need consistent TRUE/FALSE outputs. Leave higher for generative agents.
Max Response Tokens	Maximum length of the AI's response. Default: 1000.	Reduce for classification agents (a TRUE/FALSE response needs almost no tokens). Increase for agents generating longer copy.
Top P	Nucleus sampling. Controls diversity of word choices.	Rarely needs adjustment.
Top K	Available for Anthropic and Gemini. Limits token selection pool.	Rarely needs adjustment.
Frequency Penalty	Reduces repetition (OpenAI, xAI).	Useful for generative agents producing copy.
Presence Penalty	Encourages topic diversity (OpenAI, xAI).	Useful for summary agents.
Reasoning Effort	GPT-5+ models only. Controls how much the model "thinks".	Higher effort for complex analysis; lower for simple classification.

Click Reset to restore all advanced settings to defaults.

Saving and moving on

Once you've set the name, data target, provider, and model, click Continue to move to the Prompt step. You can return to the Build step at any time during creation without losing your work.