Keeping long-running agent work reliable when streams disconnect or tasks span many tool calls
Built resumable Redis-backed streaming so users can reconnect without losing agent progress
A custom production coding agent and harness with sandboxed file work, resumable streaming, tools, skills, MCP connections, and long-running context management.
Context
Structor started from the need for an agent that could do real technical work, not just chat about it. That meant handling uploaded files, editing and creating artifacts, running tool workflows, surviving connectivity blips, and keeping context under control during long tasks.
Build
I worked on the agent harness as a production system: Python backend, React/Next.js frontend, Redis-backed streaming and orchestration, sandbox containers, file sync to durable storage, MCP/tool integration, skills, patch-style editing, custom context management, and token accounting. The result is closer to a purpose-built coding environment than a generic assistant interface.
Engineering
Built resumable Redis-backed streaming so users can reconnect without losing agent progress
Designed sandboxed workspaces with file sync, S3 persistence, MCP tooling, and skill-based execution
Implemented custom context management, token counting, and orchestration around asynchronous agent work
Outcome
Stack