JAZ: Harness as a Language

An agent framework from MIT CSAIL, introduced in "Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity", built to test how far a harness that is little more than the agent loop itself can go on workloads that usually get specialized systems, such as long-horizon memory and continual self-improvement. JAZ exposes a single LLM-backed primitive, invoke, which it treats as a language construct: a function whose body an LLM writes at run time, in a Python REPL, each time it is called. The authors define the loop by two properties: the model can write arbitrary executable code, including recursive calls to invoke, and everything the model sees, both the inputs to invoke and its own interaction history, is a variable in the code environment. The design generalizes CodeAct and recursive language models (RLMs), and the second property is what separates it from code-mode agents such as smolagents and RLMs, which show the prompt and history to the model only as a list of messages rather than as data it can manipulate. Programmers add constraints and monitoring through built-in hooks for observability (logging, trajectory recording, Langfuse and Jaeger tracing), resumability (trajectory replay), resource control (shared budget pools, iteration and recursion limits, context-window warnings) and validation of return values and generated code; a hook applies either to one call or, as a context manager, to every call beneath it.

The evaluation uses prompting only, with no hand-built tools, memory or file system. On StuLife, an ordered sequence of 1,284 simulated tasks from a college semester (939 of them graded), with GPT-5.4 nano throughout, JAZ passed 69.9% of the 207 "far recall" tasks, which need information given more than 50 tasks earlier, against 61.8% for the Letta agent (the newest version of MemGPT) at less than half the cost ($18.3 against $42.1), and 32.0% for CodeAct with subagents. On AppWorld's test-challenge split, a top-level GPT-5.4 invoke acted as a meta-agent that dispatched batches of tasks to GPT-5.4 nano sub-invokes and revised prompts and skills from test feedback; it reached 74.2% task goal completion against 69.9% for the ACE self-improvement harness ($20.9 against $30.6) and 71.1% for CodeAct with subagents. Eight of the ten authors are at MIT CSAIL and two are independent researchers; Zhening Li is first author, Joshua Liu and Mateja Vukelic are marked as equal contributors, Omar Khattab is a co-author and Armando Solar-Lezama is last author. The framework (Apache 2.0, about 50 GitHub stars) has been on PyPI as jaz-lang since June 2026, and a separate MIT-licensed repository reproduces the paper's ten agent configurations.