Changelog

A workspace full of identical robot icons is hard to scan. Give each agent its own icon and color, and use that same visual cue to find it in navigation, open its chat, or select it for an automation.

You can now throw your keyboard and instaed start talking to your agent. Just hit the microphone button and it will turn your voice into text for your agent.

Agenta now works on your phone.

We redesigned session organization so you can work across more tasks in parallel without losing track of what is happening.

Sessions now live in the sidebar, so you can quickly switch between them. The new sessions view also lets you filter sessions by status, agent, or automation.

Whether you start sessions yourself or launch them automatically, it is now easier to find active work, review completed tasks, and stay organized.

When an automation stops or produces a poor result, a status does not explain why.

Automation Runs saves the full history of every run. Open a run to review the chat, give the agent input, or approve an action. You can then update the automation to improve future runs.

Changing the model, runtime, or permissions used to take you away from the chat.

Now you can type / in the message box. Change these settings or choose a skill or tool without leaving the session.

We improved security in Agenta. An agent can no longer see your API keys when it runs in a cloud sandbox. This covers both your model provider keys and the credentials on your MCP servers.

Ask an agent to print its own environment variables now, and you get a useless placeholder instead of a key.

This is on by default in Agenta Cloud, and there is nothing for you to do.

If you self-host, it applies to agents running in a Daytona sandbox, and your Daytona API key now needs permission to manage Secrets. Read the upgrade note below before you upgrade.

You can now run an agent on Codex, OpenAI's coding agent. Switching an agent over is a configuration change, and you can do it from either of the other two harnesses in the middle of a session.

Pay with an API key, or use your own ChatGPT subscription when you self-host.

Codex has the following capabilities:

• Five OpenAI models, from GPT-5.6 Sol down to the cheaper Luna, most with a context window of about a million tokens.
• Runs on the local sandbox or in a Daytona cloud sandbox.
• Follows the same tool approvals as your other agents, and answering an approval resumes the session instead of restarting it.
• Connects to your own HTTP MCP servers.
• Reads image attachments.
• Keeps every tool you already configured, with tracing and cost reporting attached as usual.

Your agents now work from a shared folder in the cloud, not from files sitting on one person's laptop.

Upload images, PDFs, spreadsheets, or a whole folder. Paste from your clipboard, drag files in, or use the attach button. Your team sees the same files, and so do your agents.

Each agent keeps its own folder that survives between conversations. Save the reference material, the approved examples, and the notes it should follow, and the agent starts every session already knowing them instead of being told again. Files you attach to a single conversation stay with that conversation.

This works on Agenta Cloud and on a self-hosted deployment.

Also in this release:

• Pi's seven built-in tools (read, bash, edit, write, grep, find, ls) are always available now. Control them from the permissions panel instead of the agent configuration.
• Agent sandboxes ship with more shell tools out of the box: ripgrep, fd, jq, unzip, zip, tree, file, process tools, and a python alias for python3.

Bug fixes:

• Approving a tool call no longer evicts a warm agent session, so resuming after an approval is fast again.
• Denying a tool call now reliably ends the turn.
• A resolved approval no longer reappears after you reload the page.
• The session Files drawer now reliably shows that session's own files.

Agenta is now a workspace for building and running agents, not only a platform for managing prompts. You build an agent by chatting with it: describe the job, connect the apps it needs, and correct it through feedback until it works the way you want.

An agent is made of instructions, tools, skills, permissions, and files. You can work with it in chat, or run it in the background on a schedule or when something happens in an app you connected. Agenta runs agents on Claude Code or Pi, records a version every time the configuration changes, and traces every run.