JIT vs rrsi
JIT-Agent: a meta-agent that writes a task-specific harness on the fly — memory, planning, action and capability modules emitted as code — and revises it from traces at test time. — versus — Google Research's Regularized Recursive Self-Improvement: evolves an agent harness (prompts, tools, memory, sub-agents) with constraints on the search so gains hold out of distribution.
Both let an LLM rewrite an agent's harness from execution traces; JIT-Agent emits a task-specific harness at test time, RRSI evolves one incumbent harness across rounds with acceptance gates.
| JIT | rrsi | |
|---|---|---|
| Stars | 533 | 1.5k |
| Forks | 58 | 141 |
| Language | Python | Python |
| License | NOASSERTION | Apache-2.0 |
| Last activity | 1 months ago | 18 days ago |
| Topics | agents, training | agents, training |
| Curated connections | 3 | 3 |
JIT — the curator's take
Research code for one sharp idea: the scaffold around an LLM is a trainable, transferable axis of intelligence, separate from the base model. Given a task spec, a protocol and a tool registry, JIT-Agent-27B emits an executable harness (four modules over HarnessFactory interfaces, so it's structured code, not a free-form program), then repairs it as feedback arrives while the generator stays frozen. Reproduce the paper with the runners and benchmark adapters. NOT a framework you build on — no licence file, conda env, benchmark-shaped — and it assumes you serve the meta model yourself (vLLM/SGLang over HTTP).
rrsi — the curator's take
Read and run it if you evolve harnesses against a benchmark and keep watching in-distribution gains evaporate — annealed edit budgets, a leakage critic, a noise-adjusted acceptance floor and a cost rule are the fix, each mapped from paper to code, with every candidate a git worktree. Not a product or a plug-in: it's paper code wired to Claude Opus on Vertex AI and three specific benchmarks (Terminal-Bench, Harvey LAB, EngDesign), and a run costs real compute.