October review: workflows that code, Miraxa builds you can test drive, and Sonnet 5.5

Route planning, forecasting, a website audit that runs itself every Monday: jobs like these need real computing power, and now a workflow has it. The Sandbox step gives each run a Linux machine with internet access, Python and the tools of the trade, and it installs the packages the job calls for.

Write a Bash or Python script for full control, or switch to agent mode and simply describe the goal: the agent works out the commands and runs them until the job is done. Files flow in from earlier steps and out to later ones, so one step can build a report and the next can save it or send it on.

Need a password or API key that no connector covers? Store it as a Secret connection and hand it only to the steps that need it; its value never shows in a step's output or logs. Scripts are billed for runtime only, and agent mode also uses model credits.

Because a sandbox step can reach the internet, give it only the inputs and connections it needs.

Tell Miraxa what you want automated, from any page, and it plans the workflow with you, opens the editor and builds it step by step on your canvas.

Every step is checked against your real connectors as it goes in, from the tools it calls to the fields it needs and how data flows from one step to the next, so what lands on your canvas fits together. A short build report under the message sums up what was built, which connections it uses, and anything left for you to fill in.

Then take it for a test drive. A test run uses your real data for the steps that read, such as fetching orders or looking up a customer, and simulates the ones that would change something, so you can watch the whole workflow at work without sending an email or touching a record.

Any action step can now be set to continue on error. If it still fails after its retries, the workflow keeps going, and the next steps can see what went wrong in a variable named after the step, such as orders_error.

Pair it with a condition and you have a fallback built right into the workflow: try one carrier and switch to another if it is unavailable, or check a server and alert someone only when it is down. Approvals and governance rules apply just as before.

Anthropic's newest Sonnet brings near-Opus reasoning at Sonnet speed, for less than Sonnet 5. It is a strong pick for busy agent steps where Opus is more than you need, and it is now the default for new workflows on the free plan.

Haiku 5.5 is the fastest and lowest cost Claude, made for classification, extraction, routing and high-volume runs. A model call whose prompt is over 100K tokens costs more, so keep it on the short steps.

Both are ready in your workflow steps, and Miraxa stays on Opus 5.5. Sonnet 5 and Haiku 4.5 remain available as previous releases.

Google's newest Gemini is in your workflow steps, for complex reasoning, coding and long multi-step agent work. It costs the same as Claude Opus 5.5 and supersedes Gemini 3.1 Pro, which stays selectable for now.

A reminder from last month: Google retired Gemini 2.5 Pro and Gemini 2.5 Flash on 20 October, so a step still set to either no longer runs. Switch it to Gemini 3.8 Flash or Gemini 3.5 Flash-Lite.

Everything here is live now. Think of the one job you still do by hand because no tool could do it, and ask Miraxa to build it with a Sandbox step.