Designing a coding agent loop without KV cache using Jev
A working note based on design notes from Jev's founder stores tool calls, diffs and AGENTS.md as typed, addressable chunks, with Jev deciding each turn which chunks to show and which model and tool to use.
Coding agents are a while loop around a model and a handful of tools. That's it. That's the whole secret. And almost nobody has rethought that loop since it was invented. Someone just compiled a working note built on design notes from Diogo Almeida, the founder behind Jev, and it asks one question that breaks everything: How would you build a coding agent if models had NO KV cache? Answer it honestly and 6 things every agent does today start looking wrong. Routing that loses money. Tool calling that floods the context. Compaction that compresses blind. Sub agents that barely ever fire. Restarts that throw away state. And the endless "batteries included" debate. The alternative is a harness built around explicit state. Every tool call, diff, user turn and AGENTS.md file stored as typed, addressable chunks. Then Jev answers the small questions every single turn. Which chunks do I show? Reuse the cache or not? Which model gets this? Which tool? Is this command allowed to run? Frontier model only when it's needed. Sub agents, cheap open models and background reviewers for everything else. Hundreds of tools, shown snippet first, full schema only on demand. This is the clearest picture I've seen of where coding agents go next. Save this. Read it twice. follow @cyrilXBT
This post quotes another post: x.com/1373665408193036293/status/210151513079829…