What if the bottleneck for AI agents isn't the model, but the instructions you give it?
A new open-source project called "Most Capable Agent System Prompt" takes that premise seriously. Rather than a traditional software tool, it's a massive, structured system prompt designed to turn AI coding assistants into more autonomous agents. It works across Claude Code, Cursor, OpenAI Codex, and other agent platforms.
The approach treats agent design as systems engineering. Instead of vague instructions like "be helpful," the prompt defines explicit task queues, approval gates, failure handling, and self-improvement loops. It separates low-risk actions (reading files, running tests) from high-risk ones (deploying code, modifying databases) and requires explicit approval before the dangerous stuff happens.
Some of the more interesting design choices:
- Five active queues (now, next, blocked, improve, recurring) that keep the agent moving between tasks without stalling
- File-first state management, meaning a project should be continuable from its folder alone, with chat history treated as optional
- Specialized harnesses for different domains like contracts, finance, and research, each with structured phases and outputs
- Reliability math: the prompt explicitly warns that a workflow with 90% step reliability still fails too often for real use and pushes for deterministic validation at each step
The project is runtime-agnostic, meaning it doesn't lock you into a specific platform. Whether you're using Claude Code or Cursor, the same architectural ideas apply.
This is still a prompt, not production software. There are no benchmarks proving it outperforms simpler setups, and the ambitious scope means most users will need to adapt it heavily. But as a reference architecture for how to think about agent reliability and autonomy levels, it's a useful read for anyone building agentic workflows.