Configure

Agents & models

TL;DR

Agent Planners ships an orchestrator that plans plus six specialist agents — Analytics, Optimization, Creative, Research, Safety and Report — that execute. Every agent runs on whatever model you choose through one built-in AI Gateway, switchable any time on the Agents page.

The agent team

AgentRole
OrchestratorDecomposes your goal into a structured, observable plan and delegates steps.
AnalyticsRead-only analysis across Google Ads, GA4, BigQuery and a Python sandbox.
OptimizationCampaign, keyword, budget, PMax and RSA mutations (approval-gated).
CreativeAd copy, asset descriptions and video scripts.
ResearchWeb research, competitor teardowns and SERP analysis.
SafetyRisk checks, stale-state validation and policy verdicts before writes.
ReportSummaries and the final readable report.

Choosing a model per agent

Open the Agents page to see each agent’s description, its tools, and a model dropdown. Set a different model for each agent through the built-in AI Gateway — there’s nothing to install and one key covers every provider.

By default the orchestrator runs on a quality-first planner and the specialists run on a fast, low-cost model, which keeps most of your credit spend cheap. You can override either per organization.

  • OpenAI GPT-5.x family (incl. Pro and mini tiers).
  • Anthropic Claude (Opus, Sonnet, Haiku).
  • Google Gemini (Pro, Flash, Flash-Lite).
  • DeepSeek, Qwen, GLM, Mistral, Llama, Kimi and more.
Reasoning models ignore the temperature setting; for other models you can tune temperature per agent (default 0.2 for deterministic, structured output).

Picking the right tier

  • Keep the orchestrator on a strong model — plan quality drives the whole run.
  • Leave specialists on a fast/cheap model for most work; bump a specific one up for hard creative or research tasks.
  • Premium reasoning models cost the most credits per token — reserve them for where they pay off.

Frequently asked questions

Can I use my own model API key?
Yes, two ways. Model access runs through the built-in AI Gateway by default, so you don’t need a separate provider key — pick any supported model per agent and usage is metered in credits. Or connect your own OpenAI-compatible endpoint via Bring your own model and pay your provider directly instead.
Does changing a model affect safety?
No. The approval gate, audit log and stale-state checks are independent of which model an agent runs on.

Related