Changelog

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.

We have rebuilt the playground from scratch to be the main workspace for your AI applications.

We have added two major features:

1. You can attach evaluators to the playground to score outputs as you edit.

2. You can load test sets. This allows you to work on your evals and add test cases directly from the playground.

Agenta now has a dark theme across the whole app. Switch between light, dark, and system from the top bar, next to the version label. Pick system and Agenta follows your operating system. Your choice is saved for next time.

The most useful thing you can do when building an LLM app is read your traces. You find the failures, label what went wrong, and turn the worst ones into test cases. That loop used to happen in spreadsheets. Annotation queues bring it into Agenta.

Build a queue from traces or test set rows, attach a scoring schema (ratings, dropdowns, rubrics, or free text), and route it to reviewers. When the queue is done, export it as a labeled test set. The annotations come along as columns, so the work feeds straight into your evaluators.

All application invocation endpoints (/generate, /generate_deployed, /test, /run) are now unified into a single POST /services/{service}/v0/invoke endpoint. The new format uses structured references for targeting applications and environments, and returns outputs under data.outputs with trace_id and span_id for observability.

The old endpoints remain available temporarily via an adapter. See the full guide for before/after examples.

You can now trigger automations when a prompt deployment happens in Agenta. Send the event to any HTTPS endpoint, or call GitHub directly with repository_dispatch or workflow_dispatch.

This makes it easier to connect prompt deployments to CI, repository sync jobs, and pull request workflows. If your GitHub workflow needs the latest prompt content, fetch it from Agenta during the run and commit the result back to your repo.

Learn more: Webhooks | GitHub Automations

You can now connect 150+ external tools to your prompts directly from the playground. Browse integrations like Gmail, Slack, Notion, Google Sheets, and GitHub. Authenticate with OAuth, attach tool actions to your prompt, and execute tool calls with one click. Use Google Sheets or Notion as data sources for RAG, send emails from your prompt, or automate developer workflows.