Flue debuts programmable agent harness

This feels less like another agent SDK and more like a runtime opinionated around real workflows, which is the right direction if agents are supposed to do durable work instead of chat demos.

• –The sandbox-first design is the differentiator: Node, Workers, GitHub Actions, and hosted sandboxes make it easier to run agents where the work actually happens
• –Skills, sessions, and explicit shell/tool boundaries give Flue more structure than raw prompt loops, which should help with repeatability and debugging
• –The positioning against tools like Dosu, Greptile, and CodeRabbit suggests this is aimed at production agent apps, especially code and ops workflows
• –TypeScript-first matters here because most agent plumbing in this ecosystem still lives in JS/TS, and that lowers adoption friction
• –There is no visible Product Hunt listing, so this is probably surfacing through X and the project site rather than a PH launch