agentic-context-engine vs JIT
Learning loop for any agent: reflect on failures, distill strategies into a Skillbook, inject them next run — 2x consistency on Tau2, 49% token cuts. LiteLLM-based, 100+ providers. — versus — 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.
Both learn from failures between runs. ACE distils strategies into a Skillbook injected into the next prompt; JIT-Agent rewrites the scaffold itself — memory, planning, action modules — from the traces.
| agentic-context-engine | JIT | |
|---|---|---|
| Stars | 2.6k | 381 |
| Forks | 306 | 43 |
| Language | Python | Python |
| License | Apache-2.0 | NOASSERTION |
| Last activity | 7 days ago | 9 days ago |
| Topics | memory, agents | agents, training |
| Curated connections | 6 | 2 |
agentic-context-engine — the curator's take
The in-process answer to 'my agent repeats the same mistakes': wrap your agent, feed it corrections, and ACE extracts reusable strategies it injects on later runs — no fine-tuning, no reward signals, and the numbers are concrete (2x pass^4 on Tau2, ~$1.50 to learn its way through a 14k-line translation). Pick it over a memory *service* when you want the learning inside your Python process rather than behind an HTTP API. NOT magic memory: strategies come from explicit feedback loops you wire up, quality follows the judge model, and the open-source engine is the on-ramp to the hosted Kayba product — check where the managed line lands before betting infra on it.
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).