Article
Jev team's public Coding Agent design draft: deciding what the model should see anew each round
In a Beating AI news brief, TypeSafe founder Diogo Almeida published a Coding Agent design draft and invited the community to experiment with it directly. He said frankly that the team probably would not have time to implement all of these ideas themselves.
The draft's core change concerns context. Current Coding Agents typically continue carrying previous conversations, tool results, and code state, compressing or restarting only when the context becomes too long. TypeSafe wants the Agent to reassess each round: whether to continue using the existing cache or select a more suitable context anew; which content to retain in full, which to keep only as summaries, and which to discard entirely.
Almeida believes that many existing Agent designs have become constrained by the KV Cache. For example, when a task is first assigned to a cheaper model and then switched back to a more capable model, the more capable model may have to reread the long context, ultimately costing more; tool definitions that remain embedded in the system prompt also continue taking up space.
If context can be dynamically reorganized, model routing, Sub-agent, MCP, and AGENTS.md can also be loaded as needed. Simple tasks can be assigned to cheaper models, only the required tools can be loaded, and each Sub-agent can receive only the state relevant to its own task. Almeida provisionally calls this approach “Meta-attention.”