# Coding agents on Blaxel

Getting an agent into a sandbox is the start. Keeping the whole workflow useful is the interesting part.

Author: Michael Stolarz
Published: 2026-09-06
Updated: 2026-09-06
Canonical: https://michaelstolarz.com/projects/agents-on-blaxel/
Project status: Ongoing

At Blaxel, I work on developer and agent experience. Here are two public
pieces of that work.

## Amp: run, reconnect, continue

I wrote the Amp sandbox tutorial around a complete path: create a sandbox,
run a coding task, read the structured result, and continue the same thread
after reconnecting.

The reconnection matters. A useful example should keep working beyond the
first command. It should explain where the state lives and how to find it again.

The guide includes TypeScript and Python, streaming output, terminal use,
and cleanup.

- [Run Amp on Blaxel](https://docs.blaxel.ai/Tutorials/Amp)
- [The change and its verification](https://github.com/blaxel-ai/docs/pull/727)

## Keeping a harness integration working

I also worked on the Blaxel integration for DeepSeek Harness. When the harness
removed its API Proxy service, the web entry could no longer start. The update
connects it to the owning Host services directly and records the compatible
versions.

That is a less visible part of an integration: following changes on both sides
and keeping the documented path usable.

- [Explore the integration](https://github.com/blaxel-ai/deepseek-harness-blaxel-sandbox)
- [The Host API migration](https://github.com/blaxel-ai/deepseek-harness-blaxel-sandbox/pull/9)

## Related

- [SystemSculpt](/projects/systemsculpt/) — An AI workspace inside Obsidian, close to the notes and files you already use.