Models and instructions

Every agent pairs an AI model with written instructions. This page covers picking the model, writing instructions that hold up, and the advanced settings for reasoning effort and context management.

Choosing a model

The model picker groups models by provider (OpenAI, Azure OpenAI, Anthropic, Google) and highlights quick picks:

• Recommended - best balance of speed and quality for most agents
• Smartest - deepest reasoning for hard analytical work
• Fast & Cheap - high-volume, low-latency tasks

Start with the recommended model. Move up only if the agent’s reasoning falls short; move down if you’re running high-volume tasks where speed and credits matter more than depth.

Writing instructions

Instructions are the agent’s system prompt (up to 100,000 characters). Good instructions typically cover:

• Role - who the agent is and who it works for
• Task - what it should do, step by step where order matters
• Rules - what it must always or never do
• Output - the format and tone of a good answer
• Context - company background, product names, terminology

Advanced settings

The Advanced panel in the Instructions section fine-tunes how the model runs. It holds two controls.

Reasoning effort sets how hard reasoning-capable models think before they answer (models without a reasoning mode ignore it). Raise it when answers come back shallow or skip steps; lower it for speed and lower credit use on simple, high-volume work.

Setting	When to use it
None	Take the model’s default behavior
Light	Quick analysis where speed matters
Medium	Balanced reasoning for everyday tasks
High	Deep thinking for hard problems, at the cost of speed

Context management decides what happens when a conversation grows long enough to fill the model’s context window.

• Auto-summarize long context is on by default. The agent compacts earlier turns automatically as the conversation nears the limit, so it never runs out of room. Leave it on unless you need the full history kept word for word.
• Summarize threshold sets how many tokens the conversation can reach before it compacts, anywhere from 50k to 400k, or leave it on Use default. Lower it to summarize sooner and keep replies fast; raise it to hold more detail before compacting.

With auto-summarize off, the agent stops compacting on its own. A Summarize button appears in chat so you can compact manually when the context ring runs high.