Organize your agents, get notified, and enjoy a cleaner chat

March 5 – March 12, 2026

---

Organize agents into folders

You can now create folders in your library to keep your agents organized. Drag and drop agents between folders, rename them, and color-code them to match your workflow. Whether you have five agents or fifty, everything stays tidy. ↗

Organizing agents into color-coded folders in the AutoGPT library

Get notified when background chats finish

AutoPilot now sends you a notification when a background chat completes. No more switching tabs to check — you'll see a toast alert the moment your agent finishes its work, so you can jump right back into the results. ↗

AutoPilot notification appearing when a background chat finishes its work

Cleaner reasoning & chat experience

We've redesigned how reasoning steps and tool calls appear in chat. Intermediate thinking is now collapsed by default, keeping your conversation clean and easy to read. Expand any step if you want the full details. ↗

Collapsed reasoning steps keeping the AutoPilot conversation clean and focused

Give feedback and share outputs instantly

Every response now comes with action buttons to copy, upvote, downvote, or listen to the output. Give quick feedback on any response so AutoPilot keeps getting better, and copy or share results in one tap. ↗

Output action buttons for copying, rating, and sharing AutoPilot responses

---

Improvements

• Credential selector redesign ↗ — Completely rebuilt the credential picker for a smoother, faster experience when connecting your accounts.

• Pinned tool cards ↗ — Interactive tool cards now stay pinned outside collapsed reasoning, so results are always visible without expanding steps.

• Workspace file uploads ↗ — Upload files directly to your workspace and reference them across agents and conversations.

• Multimodal vision support ↗ — AutoPilot can now analyze images and PDFs you share in chat, powered by the latest vision models.

• Per-turn summary stats ↗ — Each agent turn now shows a work-done summary so you can see exactly what happened at a glance.

• Text-to-speech voice selection ↗ — Choose from multiple AI voices when listening to agent responses, with improved quality to avoid robotic-sounding output.

• Batch undo in agent builder ↗ — Undo multiple changes at once when editing agents in the visual builder.

Fixes

• Session message preservation ↗ — Title updates no longer overwrite session messages, keeping your full conversation intact.

• Transcript reliability ↗ — Refactored transcripts to an atomic full-context model, preventing data loss during complex multi-step runs.

• File download integrity ↗ — Workspace file downloads are now buffered to prevent truncation of large files.

• Node layout handling ↗ — Improved how the builder handles discriminated unions and node positioning when editing agents.

• Triggered agent runs ↗ — Manual run attempts for triggered agents are now handled gracefully instead of failing silently.

• Password reset flow ↗ — Supabase error details now pass through correctly during password reset, giving clearer feedback when something goes wrong.

Under the hood

• New models added — Claude Sonnet 4.6, Grok 3, Mistral Large, Medium, Small & Codestral, Perplexity Sonar Reasoning Pro, Google Gemini 3.1 Flash & Pro, Phi-4, and Cohere Command A are now available.

• File-ref protocol ↗ — Internal file references now use a consistent protocol, making cross-block file passing more reliable.

• Langfuse tracing ↗ — Better observability for agent execution with enhanced trace correlation and feedback loops.

• Sandbox auto-pause ↗ — Code execution sandboxes now automatically pause when idle, eliminating unnecessary billing.

• Claude Code auth ↗ — Added subscription authentication support for Claude Code integration in SDK mode.