Jev Team Publishes Draft of Coding Agent Design: The model independently determines what to focus on each round.

Beating AI Express (Dongcha) reports that TypeSafe founder Diogo Almeida has released a draft design for a coding agent, inviting the community to test it directly. He openly stated that his team likely won’t have the time to implement all these ideas themselves. The core change in this draft is context management. Current coding agents typically retain all previous chat history, tool outputs, and code state; when the context grows too long, they either compress it or reset entirely. TypeSafe aims to let the agent re-evaluate each round: whether to keep using the existing cache, or select a more appropriate context; which content to retain in full, which to summarize, and which to discard entirely. Almeida argues that many existing agent designs are constrained by KV cache. For example, when a task is first routed to a cheaper model and then switched back to a more powerful model, the latter often has to re-read the long context, resulting in higher overall costs. Tool definitions, permanently embedded in system prompts, also take up constant space. If context can be dynamically reorganized, model routing, sub-agents, MCP, and AGENTS.md can also be loaded on demand. Simple tasks are assigned to cheaper models, required tools are loaded as needed, and sub-agents only access state relevant to their own tasks. Almeida has temporarily named this approach "Meta-attention".