Changelog

If your agents use a Stable Claude model, they moved to the newest Claude models over the past few weeks without you changing a setting. And you can now open a real conversation and see what the agent did.

🧠 New models, already running

Claude 5.5 is the new Stable. Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 now back the Claude Stable rows in the model picker. Agents on a Stable Claude row are already on 5.5. Agents pinned to a specific version stay where you put them. How Stable and pinned versions work.

GPT-6.1 Sol. On Claude Agent SDK agents, the GPT Sol Stable row now runs GPT-6.1 Sol.

Gemini 3.8 Flash. Claude Agent SDK agents offer Gemini 3.8 Flash in place of 3.7 Flash. Agents that were on 3.7 Flash were moved to 3.8 Flash for you.

Claude Fable 5.1. Now in the model picker for agents that run on your own Anthropic API key.

Try a model before you switch. In Improve & Try, pick a different model for a single turn. The agent's saved model, and every channel it serves, stays the same.

Check yours: open an agent, go to Settings, and look at Model Selection.

🔍 See your agents' conversations in Activity

Each agent has a new Insights → Activity tab: a feed of real conversations across your channels. Open one to see the trace behind the reply, step by step, including tool calls. Learn more.

⚡ Also shipped

A faster editor. Instructions is now an outline with an inline editor and Cmd/Ctrl+S. Tools is a list. Open a tool to switch individual actions on or off. Instructions, Tools.

Easier to find your way around. The Integrations page groups what you have connected into Channels, Knowledge and Tools, with a catalog you can browse by category. Settings is grouped into Account, Organization, Security and Billing.

More Slack options on the Claude Agent SDK runtime. Reply without @mention, and Trigger by keyword for top-level posts. Runtimes.

Another Slack workspace. Add a second workspace from the Slack integrations page, with the shared @Runbear bot or your own custom bot.

Your own Slack welcome. Custom Slack bots can set the greeting and options in Slack's assistant panel, in any language.

Clearer knowledge sync. Google Drive and SharePoint files we can't index now show as skipped, with the reason, and .md, .json and code files inside a selected Drive or SharePoint folder now sync. Disconnecting a source erases its indexed content.

Take over a knowledge source. If you don't own a Notion, SharePoint or Google Drive source, connect your own account to take it over. What is already synced stays.

Spreadsheets you can find by name. CSV and Excel files synced from now on also match a search for their filename, not just their contents.

A disconnected tool no longer ends the reply. If a tool needs someone to sign in, the agent still delivers its answer.

Connectors. A native Algolia connector, an API key option for Linear, and full action lists for 13 more connectors so you can switch off individual actions. Agents deployed to Zendesk now see the files attached to a new ticket.

Smaller touches. Browser tabs are titled after the page you are on, and knowledge source sheets save when you click Done and discard on Cancel.

You already wrote your agent's instructions — they're in your repo. Now you can ship them instead of re-authoring them in the dashboard.

🖥️ Ship your project as a teammate

The Runbear MCP server (early access) lets Claude Code, Claude Desktop and Cursor manage your agents. The runbear:deploy skill packs a local project and deploys it.

Deploy your project
/plugin marketplace add runbear-io/skills
/plugin install runbear@runbear-skills

/runbear:deploy --new "Triage"

The agent runs your files. On its next activation it loads the same CLAUDE.md, skills and code you have locally. Slack is a second ask in the same session.

⚡ Also shipped

Six managed MCPs. Airtable, ClickUp, Apollo.io, Todoist, Salesforce and Amplitude — pick the tile and authorize, no server to stand up.

⌘K in the dashboard. Jump straight to an agent, a page or an action.

Gemini 3.7 Flash. Alongside 3.6 Flash, 3.5 Flash and Flash Lite, and 3.1 Pro. The 2.x generation is retired.

More model control. The full OpenAI catalog, reasoning effort and verbosity, and stable or pinned Anthropic versions with thinking effort.

Always-on memory compaction. Long threads keep what matters without the context growing unbounded.

Setup lives in Agent Skills.

Hand a Runbear agent a real task and it now runs the whole thing to the finish, right in the thread, instead of stopping at the first reply. It can also set itself up straight from chat, and it gets sharper from the feedback you give it. Here's what's new.

🤖 Introducing the Advanced agent (Beta)

This release's headline is a brand-new agent type. Give an Advanced agent a real task and it works through every step to a finished result, right in the thread, instead of stopping at the first reply.

• Does the whole task, start to finish. Hand it a real task and it runs every step to a finished result, then drops the file back into the thread, instead of stopping at the first reply.
• Works in a real environment. It has its own file system, so it can pull in the tools it needs, run code, and produce actual files like a PDF report, not just text.
• Reuses your skills, remembers your context. Teach it a procedure once and it repeats it the same way every time, and it holds on to details from earlier conversations so nobody has to re-explain.

Turn it on. The Advanced agent adds skills, a real file system, and adaptive learning. Creating one takes a few clicks.

• Open Agents and click Add Agent.
• Name it, then choose Advanced. Standard is the fast default; Advanced is the one with skills, a file system, and adaptive learning.
• Click Create Agent. Your Advanced agent is ready.

🛠️ Build your agent right from chat

Setting an agent up used to mean clicking through settings. Now you tell it what you want in plain language and it wires itself up.

• Connect the tools it needs. Ask for something that needs an app the agent isn't hooked up to yet, and it finds the right one and connects it on the spot.
• Schedule recurring work. Say “every Monday at 9am, post a summary” and the agent sets up the scheduled job itself.
• Set up triggers. Tell it to act when something happens, like a new ticket or a form submission, and it wires up the trigger for you.
• You stay in control. Agents only reach for tools your workspace allows, and any job or trigger they set up follows your existing permissions and can be reviewed or turned off at any time.

🎓 Give feedback, teach your agents

Your thumbs-up and thumbs-down stop vanishing into a void. Now they feed straight back into how your agent works.

• Rate any reply. A quick 👍 or 👎 in Slack or Teams tells your agent what good looks like.
• Suggestions you can actually use. Runbear reviews your feedback nightly and drafts specific edits to your agent's instructions. You see exactly what would change and apply it after a quick look.

These upgrades are rolling out over the coming weeks, so some may not be switched on in your workspace yet. Your agents pick them up automatically as the rollout reaches you.

That's the update. Most of it came straight from your feedback, so keep it coming.

We've been shipping fast. Here's a roundup of what's new in Runbear since our last update.

New

• Triggers & scheduled automations. Your agents can now act on their own. Set up scheduled jobs (with a Run Now button and an active/inactive toggle) and event triggers that fire on real-world events, with a guided setup wizard, LLM-based pre-filtering, and automatic reconnect prompts when an integration needs re-auth.
• AI-assisted agent builder & editor. Build an agent just by chatting, and refine its Instructions and Contexts with an AI chat panel. We also auto-generate agent descriptions and redesigned the navigation and Overview so everything is easier to find.
• Slack-first onboarding. Getting started is now a single Slack-first flow with a persistent "Get Started" checklist and a friendlier welcome that walks you to your first working agent.
• More integrations. Official Atlassian Jira and Confluence support, plus Granola, HiBob, Tiro, Apify, Google Cloud, native Gmail and Calendar, and custom MCP servers for personal agents.
• The latest models. Claude Opus 4.7 (BYOK) with adaptive thinking, Claude Sonnet 4.6, and Google Gemini via Vertex AI. (Claude 3.7 Sonnet has reached end of life.)

Improved

• A smarter Inbox Agent. Sharper Slack and Gmail draft replies that match your voice and tone, missed-message reminders so nothing slips, and less-intrusive insight notifications.
• Slack experience. Native streaming replies with a live status pill and plan card, interactive choice buttons, response feedback buttons, and proper tables in Slack, plus a long list of rendering and reliability fixes.
• Knowledge base quality & control. Exclude individual URLs from website sync, add an entire Google Drive at once, multi-tab Google Docs support, better Notion parsing, per-agent retrieval tuning, and stronger grounding to reduce hallucinations.
• Reliability. More resilient Microsoft Teams delivery, graceful handling of broken integrations, and automatic model fallback to keep responses flowing.

✨ What’s new

• Live in under 5 minutes: Connect your inbox and start immediately. No complicated setup. No workflow cha

• nges.
• Activates the second an email arrives: As soon as a client email hits your inbox, Inbox Agent pre-populates automatically.

• Context, already assembled: It pulls the relevant account story from the tools you already use, including CRM, calendar, docs, and past threads, and puts it where you need it.
• What you need to know, summarized: Key updates, decisions, open questions, and risks are condensed into a short, scannable summary.

• Next steps you can click: Suggested actions like update CRM, loop in a teammate, or create a task, all clickable directly from the inbox.
• Replies done for you: A clear, confident reply draft appears instantly. Send as-is or make a quick edit.

• [Coming soon] One system across tools: Works alongside Runbear’s AI Teammates in Slack, so internal execution keeps pace with external conversations.

🔑 Why it matters

• Prep work disappears: No tab switching. No searching history. No guessing what changed.
• Replies get dramatically faster: Teams testing Inbox Agent are replying up to 10× faster, largely because context and actions are already prepared.
• You sound sharp by default: Every response reflects real account context, not memory or last-minute reconstruction.
• Better timing, better outcomes: Upsell moments, renewal risk, and follow-ups surface while the thread is active, not after it’s cold.

🚥 Beta signal

• Already in use by early-access teams managing high-volume client inboxes.
• Built for seat-based rollout and enterprise use from day one.

🚀 Try it now

If your inbox drives deals, renewals, and customer relationships, Inbox Agent is built for you.

👉 Try it now at app.runbear.io/inbox.

🗓️ Have questions? Book a demo to see it working on real emails and talk through a team rollout.

🦃 Happy Thanksgiving! We’re grateful for your support and excited to share a few new updates that just landed 🚀

🆕 Message Insights

We’ve added a new way to understand what users are actually asking your AI agents — without digging through raw message logs.

🔍 Grouped User Questions

Instead of showing every message individually, we now automatically group similar user questions to highlight common themes. This helps you quickly spot patterns such as frequent requests, confusing areas, or recurring problems.

🧩 Improve Agents with Context

Inside each grouped topic, you can add additional context or instructions to help your AI agents respond more accurately. This makes it easier to iterate on agent quality and improve resolution rates.

📍 Where to Find It

Go to your Agent page → Messages tab (available in the new editor). You’ll find:

• Common topics
• Unresolved user questions
• Clustered messages for easier review

🔐 Privacy Controls

This feature is only available if your organization admin has enabled Message Insights, ensuring user privacy is respected.

🛠️ Better Trigger Debugging

You can now view status, messages, and execution logs for each Trigger — and even replay past events to test them more easily. Debugging made simple!

From Messages to Credits

We're moving from message-based to credit-based billing because not all AI interactions are equal. A simple knowledge base Q&A consumes far fewer resources than a complex workflow with multiple tool integrations (MCP) and document processing.

With credits, simple Q&A costs fewer credits while complex workflows cost more, reflecting actual computational cost. If you primarily use knowledge base queries, you'll handle significantly more interactions for the same price.

Introducing Pay-as-You-Go

Previously, hitting your plan's limit meant upgrading or stopping until the next cycle. Now with pay-as-you-go, set your own overage limits and automatically purchase additional credits when needed—no service interruptions, no forced upgrades for temporary usage spikes.

Flexible Agent Limits

Previously, if you needed one more agent beyond your plan's limit, you had to upgrade your entire plan. Each plan still includes a base number of agents, but now you can create additional agents with pay-as-you-go—paying only for what you use without upgrading your entire plan.

Why We Made These Changes

Your feedback was clear: message-based pricing didn't match how you wanted to use Runbear. Sophisticated workflows felt constrained by limits that didn't reflect their value, while simple Q&A users wanted more queries without constant limit concerns. Rigid plan tiers forced unnecessary upgrades for small increases in usage.

Event Trigger (Beta)

Runbear is now available everywhere! 🎉 We’re excited to introduce Event Triggers, a powerful new way to run your AI agents automatically in response to events.

Think of it as giving your agents superpowers—able to take action automatically:

• Auto-reply to emails in real time.
• Link new Zendesk tickets to the right Linear issues.
• Send meeting summaries to participants right after Zoom calls.

This is just the beginning. Unlock your imagination and let AI handle the repetitive work, so your team can focus on what matters most.

⚡ Beta availability: We’re rolling this out gradually during the beta period. Stay tuned as access expands!

New Agent Form (with Agent Context)

Now you can create agents by simply choosing your tools—no prompts or complicated setup required. Once created, you can continue to teach your agents with Agent Contexts. (Or you can teach them directly in Slack, e.g., “@Agent, remember that you can find revenue from here: https//docs.google.com/...”)

Prefer more control? No worries! You can always switch back to the classic editor for full customization.

Enjoy a smoother, faster way to bring your AI agents to life, without the hassle.

Other changes

• Official MCP Integrations: we’ve added official support for Notion and Hubspot MCP. This brings smoother usability and an overall better experience.
• Integration Guidance: Runbear can now recommend which integrations you’ll need when your request depends on tools that aren’t yet connected, making it easier to get set up without guesswork.

Better Smart Reply

Smart Reply just got smarter! 🎉 Runbear AI Agents now proactively offer relevant information when they find it—no extra searching needed. This way, your teammates get answers faster, without interruptions.

Other changes

• 🌐 OpenAI Web Search Update: We’ve upgraded from Bing Search to Serper API, ensuring continued support after Bing’s retirement (Aug 11, 2025). Serper provides Google-powered, AI-optimized results for more accurate and reliable answers.
• 📂 Knowledge Base Improvements: Easier, faster, and more reliable

🧠 GPT-5 Is Here Now at Runbear

Runbear now supports GPT-5 — OpenAI’s most advanced model yet, combining faster, smarter responses with built-in thinking for complex reasoning. Power your Runbear Agents with GPT-5! 👉 Learn more about GPT-5 in our deep dive and see its performance in our benchmark post.

🚀 Runbear OpenAI Agent Now Available via MCP

Your Runbear OpenAI Agent now supports MCP (Model Context Protocol), enabling seamless connections to external tools and data sources. — making your OpenAI-powered Agent smarter.

🔍 Smarter Task Analysis — From Slack Conversations to Ready-to-Use Agents

Task Analysis has been upgraded to be smarter and more actionable.

It now goes beyond detecting repetitive work — analyzing the context from your Slack conversations to understand your tasks and offer ready-to-use Agents.

Other changes

• Trigger Agents in Zendesk with Specified Keywords: Call an Agent by typing a preset trigger word in an internal comment
• Smart Reply Prompt: Include sample questions as prompts so Smart Reply is triggered more accurately when users want it
• Custom MCP with Bearer Authentication: Now you can securely connect to MCP servers that use Bearer token authentication, enabling safe integration with token-based MCP servers.