Running a personal AI agent for months: architecture, failures, and open questions
Powermadlux · reddit · 2026-11-12
An infrastructure engineer shares months of experience running a persistent personal agent 'Albert' used daily across software projects, research, writing, and music.
Key architecture:
- Context lives in Markdown alongside projects: session boot file, project state, decisions, changelogs, Git history
- Fresh sessions pick up projects without re-explaining background
- Agent handles project files, tools, and authenticated web browsing; human reviews before publishing
Biggest pain: even documented context gets missed — the agent often fails to find the right file or reads the current version, or treats old plans as current ones, and errors compound across projects.
The author asks the community about session continuity, permission/isolation practices when the agent with local access reads untrusted web content, and what to build differently today.
More from coding & agent
- Anthropic's Lydia Hallie on Claude Code best practices: 'The modern programming language is English' — ghumare64 · 2026-11-12
- Addy Osmani's guide to picking a Claude model: start with Opus at medium effort — bibryam · 2026-11-12
- Fable 5.2 ports a bespoke CUDA voxel renderer to Vulkan, running 4K on an RTX 5090 — jongranskog · 2026-11-12
- Marketing MCP connects Claude/ChatGPT to 14 ad, GA4 and SEO platforms — philrox_ · 2026-11-12
- Vincent Koc: have your agent build clickable dummy wireframes before writing real UI — vincent_koc · 2026-11-12
- Fable 5.2 Runs 32 Agents Across 3 Parallel Projects With Zero Issues — AIFlow_ML · 2026-11-12