JAZ invoke: a minimal agent loop that beats specialized memory and self-improvement harnesses
Sep 22 Li et al.: invoke makes prompts, tools, and REPL history first-class REPL variables—StuLife far-recall 69.9% vs Letta 61.8% at ~half cost; AppWorld TGC 74.2% vs ACE 69.9% with lower meta spend.
Source: Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity
Zhening Li, Joshua Liu, and collaborators ask how much of “harness engineering” is really missing expressivity in the agent loop. JAZ centers on invoke: a language primitive whose body is written by the LLM at call time, with recursive invoke as the default subagent, and—crucially—every input plus the REPL __history__ available as variables the model can pass by reference. Hooks add budgets, validation, and observability without baking memory or ACE-style playbooks into the runtime.
Prompt-only head-to-heads are the point. On StuLife’s far-recall subset (207 tasks needing facts from >50 tasks earlier), GPT-5.4 nano JAZ invoke reaches 69.9% pass / 73.6% score versus Letta (MemGPT) 61.8% / 67.0% at roughly half the cost ($18.3 vs $42.1), and far above CodeAct+subagents (32.0%). Tail-recursive delegation keeps prior __history__ by reference across dozens of context-window handoffs. On AppWorld test-challenge continual self-improvement, a GPT-5.4 meta-invoke over GPT-5.4 nano solvers hits 74.2% TGC versus ACE on CodeAct at 69.9%, with lower total and meta cost—because the meta-agent batches, selectively reads traces, and can optimize tools as well as prompts, not only a rigid reflector→curator playbook.
Why it matters for code review
Always-on factories often grow specialized memory/self-improve layers that encode brittle assumptions. JAZ’s productive complication: sometimes the scarce asset is a minimal, maximally expressive control plane plus reviewable hooks—not another external harness module. Review overnight systems for whether subagents can return lossless history by reference, whether meta-agents can change tools as well as prompts, and whether “specialized” components are still earning their rigidity. Pair with Li’s Grow the Harness, Not the Context and Self-Harness’s held-out promotion rule.