Changelog

Agenta now decides project access in one place. Creating, renaming and deleting a project follows organization roles. An API key acts only in its own project for management actions. When you remove a member or delete a key, the change takes effect at the next request.

Automations, evaluations and approvals run as the user or key that created them. If that identity loses access, the run stops and tells you why.

Prompts and evaluators now run on the provider keys saved in the project. A project without a key for a provider gets the error "No API key found for model". Add the key in Settings > Providers & Models.

Custom providers need their own API key, except Bedrock, SageMaker and Vertex. Self-hosted Enterprise deployments can opt in to keys set on the services container.

The Files pane now previews Excel, Word and PowerPoint files, so you can open what your agent produces without leaving Agenta. You can also comment on a spot in an HTML app, document, slide or spreadsheet, and the agent receives the comment with its location.

Every file and folder in a session now has a link you can share. The chat transcript, file cards, attachments and composer have a new design. Agents open straight into the playground.

Setting up an agent is now mostly a conversation. The + on Integrations, Skills and Automations has a Create with AI option that starts the request in chat, so the agent does the setup for you.

Agents created from a template also ask for the accounts they need before their first task, instead of failing halfway through it.

You can now run agents on Claude Opus 5.5, Claude Sonnet 5.5 and xAI's Grok. Connect xAI directly from AI providers, no OpenRouter needed.

Cost tracking is also more accurate. Costs now account for cached tokens and cache writes, each model call is priced on its own, and calls we can't price are marked instead of showing $0.

You can now start from a ready-made agent instead of a blank one. Pick a template in the app or in the new marketplace, and you get a working agent with its skills, files and the connections it needs.

Built something useful? Open the Template options menu in a session and choose Save as template (.zip). The agent packages itself so others can use it.

Agents connected to Slack or Telegram can now work with those channels from any run, not just reply to a message. An automation can post its weekly report to #marketing. An agent in chat can catch up on what your team said in #support.

Agents can also share the link to their current chat from anywhere, so every post can point back to the full conversation.

You can now talk to your agent in Slack. Add it to your workspace from the agent's Publish button, then mention it in any channel or thread. It answers in the thread and keeps the conversation's context.

You can see where each request stands: the agent adds 👀 when it picks up your message and ✅ when it's done.

You can now chat with your agent in Telegram. Open the agent's Publish button, choose Telegram, and scan the QR code with your phone. That's the whole setup.

The chat keeps its memory across messages, so you can pick up a task from your phone where you left it.

Agents can now answer your customers on WhatsApp. Connect your WhatsApp business number to an agent, and it replies to every customer who writes to that number, with typing indicators, approval buttons, and support for images and documents.