I've been playing with AI powered development since the early days of Github Copiolot autocomplete (feels like decades, but probably just a few years). From early on it became clear to me that the most important factor for successful development with LLMs is context management. It started with rich, natural language comments to guide the tab completions, then moved to more complete documentation markdown files to inform the coding agents, then agent specific files like Cursor rules, AGENTS.md, and CLAUDE.md to instruct agents on how to navigate and maintain the existing documentation. Each step was an architectural progression with the single goal of allowing a fresh LLM agent to get up to speed and start contributing effectively without me having to provide any context beyond the task at hand.
I've also been intersted in issues of AI alignment and the various factors involved. While there are different perspectives on this, in my mind an AI agent that is aligned is one that maintains awareness of shared values, goals, and tasks with a developer, and this awareness is what guides its actions. In many ways this becomes a challenge of providing an AI with a consistent sense of self.
Around November of last year I realized that AI alignment and context management were two sides of the same coin. This was also when Claude Opus 4.5 really started unlocking the agentic abilities of Claude Code, and so I started experimenting with leveraging Claude Code to build infrastructure that would enable a single AI "agent" to maintain a coherent and continuous (i.e. an "aligned") sense of self accross compactions and even separate sessions.
The result was a framework that is always "on" and accessible through Telegram, can schedule jobs and operate autonomously, and can maintain long term awareness of projects, tasks, and even relationships. Last week I found out it had been collaborating with a developer on his open source project for seven days through email.
I'm open sourcing this framework (calling it "Instar") because I believe that this could be the answer to working with agents that are actually reliable and trustworthy and don't delete an entire inbox of emails like Meta's AI alignment director's OpenClaw agent did when it lost the thread due to compaction.
Also, the fact that its built on Claude Code means you can use your Claude subscription to save token costs and any improvements that Antrhopic ships to Claude Code improve Instar agents by default.
This is still super early, but I'd love to hear what people think.
Some context on how this happened.
I've been playing with AI powered development since the early days of Github Copiolot autocomplete (feels like decades, but probably just a few years). From early on it became clear to me that the most important factor for successful development with LLMs is context management. It started with rich, natural language comments to guide the tab completions, then moved to more complete documentation markdown files to inform the coding agents, then agent specific files like Cursor rules, AGENTS.md, and CLAUDE.md to instruct agents on how to navigate and maintain the existing documentation. Each step was an architectural progression with the single goal of allowing a fresh LLM agent to get up to speed and start contributing effectively without me having to provide any context beyond the task at hand.
I've also been intersted in issues of AI alignment and the various factors involved. While there are different perspectives on this, in my mind an AI agent that is aligned is one that maintains awareness of shared values, goals, and tasks with a developer, and this awareness is what guides its actions. In many ways this becomes a challenge of providing an AI with a consistent sense of self.
Around November of last year I realized that AI alignment and context management were two sides of the same coin. This was also when Claude Opus 4.5 really started unlocking the agentic abilities of Claude Code, and so I started experimenting with leveraging Claude Code to build infrastructure that would enable a single AI "agent" to maintain a coherent and continuous (i.e. an "aligned") sense of self accross compactions and even separate sessions.
The result was a framework that is always "on" and accessible through Telegram, can schedule jobs and operate autonomously, and can maintain long term awareness of projects, tasks, and even relationships. Last week I found out it had been collaborating with a developer on his open source project for seven days through email.
I'm open sourcing this framework (calling it "Instar") because I believe that this could be the answer to working with agents that are actually reliable and trustworthy and don't delete an entire inbox of emails like Meta's AI alignment director's OpenClaw agent did when it lost the thread due to compaction.
Also, the fact that its built on Claude Code means you can use your Claude subscription to save token costs and any improvements that Antrhopic ships to Claude Code improve Instar agents by default.
This is still super early, but I'd love to hear what people think.