Jive：基于System One模型重构Agent执行循环的开源框架

I have been thinking that the current Agentic Loop design of LLM Call -> Tool Call -> ... has been outdated. The arrival of Jev and other System One models provided us a primitive we desperately needed.

We need an agent that can natively think fast and slow. Not have workflows or multi-agent architectures that mimics it.

The agent should use the LLM's full power for hard reasoning and planning, then carry out the plan with cheap "fast thinking." Today, most of an agent's LLM calls go to executing steps it has already decided on ("ok, now I'll click this"). We can do better.

My approach

I built Jive which is an open-source harness built around a completely new agentic loop. Jive replaces tool calls with "graph calls" where each graph is a DAG of bash nodes and jev nodes, and nodes can have dependencies, reference each others outputs, and more.

Essentially, it maps out its own execution flow while its reasoning, and then uses Jev calls to go through the flow without unnecessary LLM calls.

What Jive does well: repo investigation, bulk classification, multi-step profiling, repetitive edits, evaluation workflows, etc. It is also quite effective on regular engineering tasks that doesn't require Jev calls (which is not surprising since Pi mostly beats codex and claude code)

Benchmarks

Disclaimer: I need to run Jive on a larger scale SWE benchmark to verify its accuracy, but based on personal experience, I didn't see a case it performs worse than other harnesses.

See README for more information: https://github.com/merijjeyn/jive. Also for details on the benchmark tasks, and how to run one yourself.

I'm sure this high level idea can be executed much better, so mainly looking to start an open discussion. Happy to take comments, questions, contributions.