Context and “memory” (not really a fan of this term and how industry uses it) are actually complex to manage for power users including humans and agents.
While it may sound counterintuitive, the agents of today aren’t truly autonomous in that you need to really guide them and plan their actions well.
I believe this is true today, and will be even more true when agents are guiding agents.
We need new infrastructure for dynamic context management.
The answer is not as simple as “hook up your agent to an MCP that pull docs from the web” … also MCP needs its own revolution. I tend to use no MCP and prefer raw agent performance.
I’m evolving the simple concepts I built in my VS Code extension to address this. Nothing public now, but I and a few others use this everyday to feed parts of large codebases into Gemini (to build plans for Claude code, other coding agents):
https://github.com/backnotprop/prompt-tower
While it may sound counterintuitive, the agents of today aren’t truly autonomous in that you need to really guide them and plan their actions well.
I believe this is true today, and will be even more true when agents are guiding agents.
We need new infrastructure for dynamic context management.
The answer is not as simple as “hook up your agent to an MCP that pull docs from the web” … also MCP needs its own revolution. I tend to use no MCP and prefer raw agent performance.
I’m evolving the simple concepts I built in my VS Code extension to address this. Nothing public now, but I and a few others use this everyday to feed parts of large codebases into Gemini (to build plans for Claude code, other coding agents): https://github.com/backnotprop/prompt-tower