← @bgSponge

Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity

Summary

Zhening Li et al. introduce JAZ, whose invoke primitive is a minimal agent loop that treats LLM-written code and recursive subagents as the harness.

  1. invoke is a language primitive, not a toolkit stack.

    The LLM writes the body of each call; recursive invoke is the default subagent path, generalizing CodeAct and recursive language models.

  2. Everything the model sees is a REPL variable.

    Named inputs, the user prompt, and interaction history are programmable objects, so agents can search prior history instead of only reading a message list.

    everything visible to the model is also a variable in the code environment

  3. StuLife far-recall favors prompt-only invoke over specialized memory.

    With GPT-5.4 nano, invoke averages about 70% pass on far-recall tasks versus Letta at about 62% and roughly half Letta's dollar cost.

    With GPT-5.4 nano, it achieves 70% on the recall-heavy subset of StuLife

  4. AppWorld self-improvement also stays inside the loop.

    A top-level GPT-5.4 invoke with nano sub-invokes averages about 74% TGC, ahead of ACE on CodeAct at about 70%, while editing prompts and skills from test feedback.

    it achieves 74%, outperforming CodeAct (68%), CodeAct+subagents (71%), and a specialized self-improvement harness, ACE

  5. Hooks and dynamic scope keep the core thin.

    Observability, budgets, validation, and ConfigOverride attach locally or via scope without replacing the invoke loop.

    We study whether a minimal harness that is little more than the agent loop itself can exhibit these capabilities.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →