Configuring LLM Models
- How to select a platform and model for your LLM agent
- What the General tab settings do
- How to fine-tune temperature, reasoning effort, token limits, and other options on the Advanced tab
When you set an agent's Agent Type to llm, AutoTalk reveals the model configuration options. The first thing you choose is the AI platform, then you configure the model and its behavior.
Selecting a platform
The Platform type field determines which AI provider powers your agent. AutoTalk supports these managed providers and custom endpoint options:
| Platform | Description |
|---|---|
| openai (default) | OpenAI models including GPT-5.5, GPT-5.4, GPT-4.1, and others. |
| deepseek | DeepSeek models including DeepSeek V4 Flash and V4 Pro. |
| gemini | Google Gemini models. The Model dropdown is the authority on which Gemini families are currently offered. |
| anthropic | Anthropic Claude models including Opus, Sonnet, and Haiku. |
| openrouter | Route requests through OpenRouter by entering a provider-prefixed model ID. |
| cloudflare | Cloudflare Workers AI models (@cf/...). See Cloudflare Workers AI for the account-ID requirement and the tool-support caveat. |
| custom | Connect to a custom or third-party API endpoint. Use this for providers not listed above, or for your own fine-tuned models served through a compatible API. |
This table deliberately does not enumerate provider model IDs. The Model dropdown is generated from AutoTalk's live model catalog, so it is always the current answer to "which models can I pick?" -- a model named in an older runbook may have been retired at the provider and removed from the catalog.
After selecting a platform, five configuration tabs appear: General, Messages, Actions, Tools, and Advanced. This page covers the General and Advanced tabs; see System Messages and Tools for the others.
General tab settings
The General tab contains the platform, model, and API token.
Model
Select the specific model to use from a dropdown list. Each model displays:
- A cost badge (Low cost, Medium, or High cost), derived from where the model's combined input + output price falls among the available models
- Pricing per 1K tokens for both input and output, so you can estimate costs
Lighter models cost less but may produce lower-quality responses; heavier models cost more but handle complex conversations better. Deprecated models remain visible for existing agents but are marked in the picker; choose a current stable model for new agents.
Token
A searchable field where you can select an existing API token or create a new one. This token authenticates your agent with the chosen AI platform. If you have not yet added an API key for the platform, you can create one directly from this field.
Note: The Token field is conditionally hidden. It appears only when you use the openrouter platform, or when you turn on the Enable BYO API key toggle — also on the General tab — for a managed provider. With BYOK off on a managed provider (openai, deepseek, gemini, anthropic), AutoTalk supplies its own platform key, so no Token field is shown.
Cloudflare Workers AI
When Platform type is cloudflare, the Model dropdown lists Cloudflare Workers AI models, whose IDs begin with @cf/.
Two things differ from the other managed providers:
- Account ID. Turning on Enable BYO API key reveals a required Cloudflare account ID field on the General tab. The account ID rides the Workers AI request URL, so your own token is unusable without it. With BYOK off, AutoTalk supplies both the key and the account.
- Tool support is not universal. Several catalogued Workers AI models do not support function calling. When you pick one, the Tools and MCP Servers tabs disappear, and the agent cannot call a tool at all. If your agent needs tools, choose a Workers AI model that declares function-calling support, or use another platform.
Advanced tab settings
Temperature, token limits, and the optional behavior toggles live on the Advanced tab, grouped into sections.
Temperature
Under the Behavior section. Controls how creative or deterministic the model's responses are. The default value is 1.
- Lower values (0.0 to 0.5): The agent gives more focused, predictable, and consistent answers. Best for factual customer support.
- Higher values (0.8 to 1.5): The agent produces more varied and creative responses. Useful for brainstorming or casual conversation, but may reduce accuracy.
Many models do not support temperature at all. The form does not go by model name — it reads the capability flag on the selected model's catalog entry and disables the temperature control whenever that model declares no temperature support. This currently covers the reasoning-oriented GPT-5 family and every current Claude model, including the Anthropic default, but the rule is the flag, not the list: if the control is greyed out, that model ignores temperature and the setting would have no effect.
Reasoning effort
Under the Behavior section, shown only when the platform is openai (stored as reasoningEffort on the agent's LLM settings). It sets how much the model reasons before it answers: default, none, low, medium or high. More effort can give better answers on multi-step work, at a higher cost and latency.
default, or leaving it empty, sends nothing, so the model runs at its own default effort -- exactly as every agent did before the setting existed.- The control is disabled for a model whose catalog entry does not take a reasoning effort (GPT-4.1 and GPT-4o, for example), the same way Temperature is disabled for models without temperature. A value kept for such a model is not sent.
- Some models accept only part of the range -- one of them refuses
none, for example. A value the model refuses is dropped and the model runs at its default, rather than the turn failing. - The API mode decides whether it applies with tools. On Responses the effort is sent as
reasoning.effortand applies with or without tools. On Chat Completions it is sent asreasoning_effort, but a few models accept tools there only with reasoning switched off, so while the agent has tools those models run atnonewhatever you choose. Switch the API mode to Responses to keep the chosen effort.
When a chosen effort is not sent, or is replaced by none, the agent's log says so at the info level ("Configured reasoning effort not sent", or "...overridden to "none" on Chat Completions with tools"). Check there if the setting seems to have no effect. On any platform other than openai the field is hidden and nothing is sent. Creating an agent on another platform stores no value, while saving an existing agent on another platform keeps the value already stored -- the save ignores the hidden field without clearing it -- so it applies again if you switch the agent back to openai.
Max iterations
Under the Behavior section. The tool-call loop is always active — there is no on/off switch for it. An LLM agent may run several tool-call rounds automatically before producing its final answer, and Max iterations is what caps how many LLM rounds a single run may take: default 15, range 1–50.
This matters for cost and latency sizing: with the default, one incoming message can cost up to 15 model rounds, not one.
When Enable structured outputs is on, the agent is single-turn — the Max iterations field is hidden and the budget is clamped to 1.
API mode
Under the Behavior section. This selector appears on every platform that has more than one wire protocol, which today means openai, anthropic and gemini. It is not openai-only, and each platform's native default is the one you want unless you have a specific reason to switch:
| Platform | Modes | Native default | What you give up by switching |
|---|---|---|---|
| openai | Chat Completions, Responses | Chat Completions | Responses adds server-side response storage and native MCP handling, and keeps a chosen reasoning effort when the agent has tools. |
| anthropic | Messages, Chat Completions | Messages | Messages is the only Anthropic wire that can cache the prompt. Switching to Chat Completions silently loses prompt caching entirely, so a large system prefix is reprocessed on every call — an invisible cost increase. |
| gemini | Interactions, Chat Completions | Interactions | Interactions is the only Gemini wire that carries thinking configuration and reports thought tokens separately. Chat Completions is the only Gemini wire that accepts tool pinning, so switch only if your agent needs to force a specific tool. |
Changing this away from the native default is a real tradeoff, not a formality. On anthropic in particular, the cost effect (loss of prompt caching) is not visible anywhere in the form.
Gemini's native Interactions wire cannot carry a structured-output schema, so enabling Enable structured outputs on a Gemini agent moves it to Chat Completions.
Max run seconds
Under the Behavior section. The wall-clock budget for a single agent run: default 300 seconds, range 30–1800. A run that exceeds it is cut off. If an agent chains several slow tool calls and stops partway through with no error visible to the customer, this is the limit to raise.
Max response tokens
Under the Token limits section. Sets the upper limit on how many tokens the model can generate in a single response. Use this to control response length and cost. If left empty, the model uses its default maximum.
Max chars per message
Under the Token limits section. Limits how many characters from the user's message are sent to the model. The default is 1024 characters (maximum 2048). Increase this if your customers tend to send longer messages and you want the agent to consider the full text; decrease it to reduce costs on verbose inputs.
Enable structured outputs
Under the Structured outputs section. When checked, the model is instructed to return responses in a structured format (such as JSON). This is useful when the agent's output is consumed by another system rather than displayed directly to a customer.
Enable context token budget
Under the Context budget section. When checked, AutoTalk manages how much conversation history is sent to the model by enforcing a token budget for context. This prevents the context window from exceeding the model's limit on long conversations and helps control costs.
Enable image processing
Under the Image processing section, available only when the platform is openai. When checked, the agent can receive and process images sent by customers (on models that support vision capabilities). This allows the agent to describe, analyze, or respond to photos and screenshots.
Image quality
Under the Image processing section: auto (default), high or low. It controls how much detail the model is asked to extract from an image, trading token cost against fidelity.
This field is only visible when all three conditions hold: the platform is openai, Enable image processing is on, and the selected model declares support for user images. If you cannot find it, check the model first.
React to file attachments
Under the Image processing section, shown on every LLM agent. Off by default.
Attachment awareness controls additional handling of captioned or bare file messages. With it off, those messages can be skipped, but transcribed audio and supported images with image processing enabled are handled before this switch and can still reach the agent. It does not discard every message containing a file.
If an attachment gets no answer, check this setting together with the channel transcription profile, image-processing setting and model vision support. Turning on attachment awareness makes additional attachment data available in the agent context.
For most customer-facing agents, start with the openai platform and pick a current stable model from the dropdown. If the temperature control is enabled for that model, 0.3 to 0.5 gives reliable answers; if it is greyed out, that model ignores temperature entirely. Leave Max chars per message at the default, keep Max iterations at 15 unless you have measured a reason to change it, and enable image processing only if your use case requires it, since it increases token usage. If your agents receive files, turn on React to file attachments for additional attachment handling; transcription and enabled vision follow their own settings.