Managing context for long-running agents, across sessions, or swarms of agents, remains an active area of research and development. This is a survey of the current landscaspe in the world of Agentic memory.

After adoping the obra/superpowers workflow, design docs became an important part of my coding process. They’re detailed and verbose - containing code snippets, specific tasks and implementation procedures, and tests - perfect for handing off to subagents for implementation. But after the feature is built, I couldn’t decide whether to include them as part of my git history or to discard them. 95% of their content will probably never be consumed by either me or LLMs ever again, but they do contain little snippets of historical context that may be important for future architectural decisions.

Managing a complex Python repository with a shared core, various plugins, and multiple applications used to be a headache involving brittle path dependencies or complex scripts. In 2025, the uv workspace has emerged as the definitive solution for these "monorepo" architectures.

Managing Context: Use Subagents and Skills toegether to enhance your workflow

I don’t usually write about security, but as a frequent user of Claude Code, I felt the recent news of the s1ngularity attack warrants special attention. Just as we all started to rely on AI CLI tools like Claude and Gemini to enhance our work, hackers just exploited them in a real supply-chain attack. It’s the first time we’ve seen this in the wild.

I used to spend my time before classes playing the original Desktop Tower Defense. It had a long loading screen, it lagged when there are too many units, and it was written in Flash. I wanted to make a modern JavaScript version that’s smoother and more optimized, but I wasn’t a web developer… until LLMs came along.