> ## Documentation Index
> Fetch the complete documentation index at: https://docs.triform.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agents

> LLM logic with prompts and tools; good patterns and observability

## Anatomy

* Prompts
  * System Prompt — Non‑negotiable instructions.
  * Messages List (toggle) — Pass conversation history as input/output.
  * User Prompt — Optional pre‑set prompt for fixed‑purpose Agents.
* Model configuration — Provider/model, temperature, top‑p, max tokens.
* Toolbox — The set of callable tools made available to the Agent.

## Available Models

Triform supports a wide range of AI models:

**Anthropic:**

* Claude Haiku 4.5 — Ultra-fast responses, cost-effective for chat
* Claude Sonnet — Balanced performance and capability
* Claude Opus — Maximum reasoning power

**Google:**

* Gemini 3 Pro Preview — Latest flagship with enhanced reasoning
* Gemini Flash — Fast responses for real-time applications

**OpenAI:**

* GPT-5.1 — Advanced reasoning and coding
* GPT-4.1 — Reliable general-purpose model

**Open Source:**

* Kimi — Default chat model with excellent context handling
* Various models via Fireworks and other providers

Models are continuously updated. Check the builder for the latest available options.

## Good patterns

* Keep System Prompts specific and outcome‑oriented.
* Start with few high‑leverage tools; expand gradually.
* Use Messages List when memory/context matters.
* Leverage enhanced tool calling for more reliable agent operations.

## Enhanced capabilities

Recent improvements to the Agent system:

* **Improved tool calling** — More reliable tool usage with better error handling
* **Execution visibility** — Real-time execution steps sent to UI for debugging
* **Better prompts** — Refined prompt system for clearer agent guidance
* **Loop support in tools** — Agent tools can now use iterative operations

## Observability

* Executions emit logs, traces, metrics for inspection.
* Use Payloads to reproduce and compare runs after prompt changes.
* Monitor execution steps in real-time through the chat interface.

> tip: Tool hygiene — Name tools by verb + object (`fetch_user`, `summarize_text`). Small, predictable tools make plans more reliable.