Agent settings
Model, reasoning effort, access control, data handling, and the rest of an agent's configuration.
On this page
The Settings tab holds the configuration that is not instructions, knowledge, or tools. Which sections appear depends on the agent's runtime.
Model#
Choose the model the agent runs on:
| Model | Best for | Relative cost |
|---|---|---|
| Claude Fable 5.1 | Long-horizon agentic work, multistep research, hardest reasoning | Highest, and billed by Anthropic rather than in credits |
| Claude Opus 5 | Complex analysis, research, nuanced writing | Highest |
| Claude Sonnet 5 | General-purpose work, coding, problem-solving | Medium |
| Claude Haiku 4.5 | Quick answers, high volume, routine operations | Lowest |
| GPT-5.6 Sol | The strongest reasoning of the GPT-5.6 tiers | Highest |
| GPT-5.6 Terra | Balanced capability across most work | Medium |
| GPT-5.6 Luna | Fast, inexpensive, latency-sensitive work | Lowest |
| Gemini 3.1 Pro | Long-context reasoning and agentic workflows | Highest Gemini cost |
| Gemini 3.8 Flash | Fast agentic workflows, coding, and high-volume work | Medium Gemini cost |
| Gemini 3.5 Flash Lite | High-volume routine and latency-sensitive work | Lowest Gemini cost |
For Claude Agent SDK agents, Runbear-provided access uses Runbear credits. If the agent uses its own API key, Claude models require an Anthropic key and Anthropic bills the usage; Gemini models require a Google AI Studio key and Google bills the usage. To switch between Claude and Gemini while using your own key, first switch the agent back to Runbear-provided access, select the model from the other family, then save the key for that family.
Claude Fable 5.1 runs only on an agent's own Anthropic key. It is not offered on Runbear-provided access, so it appears in the picker only once the agent is on its own key.
A more capable model is not automatically a better agent. If an agent mostly answers from a knowledge base, the cheapest tier usually matches the most expensive one on answer quality while consuming a fraction of the credits. See billing and credits.
Reasoning#
Agents on models that support extended reasoning expose an effort setting. Higher effort improves multi-step work and tool sequencing, and costs more credits per message. Leave it low for lookup-style agents.
Access control#
By default every member of the organization can use an agent. Switching access to restricted limits it to the members you list.
Restricting an agent controls who can use it in the dashboard and in channels. It does not retroactively hide conversations that already happened.
Data handling#
Organizations entitled to HIPAA mode can set an agent's data handling mode. In HIPAA mode the agent runs on a restricted processing path; agents left on the standard mode are unchanged.
The mode is chosen per agent, and it appears in the create dialog as well, so a new agent starts in the right mode instead of being switched afterward.
Claude Fable models are not available in HIPAA mode. An agent in HIPAA mode cannot be moved onto one, and an agent already on one cannot be switched into HIPAA mode.
Before switching modes, read the HIPAA setup guide for Enterprise eligibility, BAA requests, supported runtimes, knowledge migration, and data-handling limits.
Long-term memory#
Memory is on by default and is documented on its own page.
Response feedback#
Controls whether the agent invites reaction-based feedback on its replies. The reactions collected here are what the analytics dashboard reports as user feedback.
Tool activity#
While an agent works, it reports each tool call as it happens — a task card in
Slack, and a thread.tool_call.progress event on the API and SDK streams. The
Tool Activity section controls whether this agent does that.
The setting is on by default, and turning it off is a presentation change only: the agent still runs exactly the same tools, and the run's own trace still records every one of them, so the activity remains available to anyone reviewing the run afterwards. What stops is the live narration, on every surface at once — a reply arrives without the intermediate steps.
Turn it off when the steps are noise to the people reading, or when a custom frontend built on the API should show only the answer.
The section appears on every agent, but not every runtime narrates its tool calls in the first place — Teams and Discord replies never show them, and only some model families report them at all. On those the switch has nothing to hide, which is why it stays available rather than disappearing.
Choice buttons#
Rather than ending a turn with a typed-out question, an agent can offer the answers as clickable options — "Approve" and "Reject", or a short list of the choices it was about to describe. The Choice Buttons section controls whether this agent may do that.
The setting is on by default, and turning it off withdraws the buttons on every surface the agent answers on, not just one. Turn it off when the agent should carry out what your instructions authorize without stopping to ask.
The section appears on every agent, but not every runtime can offer buttons. If the setting is on and buttons never appear anywhere, the agent's model family is the thing to check with your Runbear contact before treating it as a fault.
Clicking an option runs the choice as the person's next turn, so it is billed like any other message. The conversation is never blocked on a click — answering in your own words instead is always valid.
Where the buttons appear, and how long they stay clickable, depends on the channel. Slack renders them as message buttons and deliberately leaves them in place after a selection, so someone can try another option later; each of those clicks runs as its own turn. Rendering them in the Web SDK is a separate closed beta — Runbear enables it per organization, and it is then configured per agent, so the switch above being on is necessary but not sufficient there — and there the buttons go inert once a later message answers them.
In that beta the switch reaches further than its name: it also governs the card, a formatted block of details the agent displays without asking anything. Turning choice buttons off withholds both.
How often the agent offers them there is a separate per-agent dial, set over the REST API rather than from this page: see the Response Components API.