JAZ: Harness as a Language, a Minimalist Agent Framework with Maximal Expressivity

Source: arXiv 2609.26891 (Zhening Li, Joshua Liu, Mateja Vukelic, Nicole Shen, Supriya Lall, Alex Zhang, Omar Khattab, Armando Solar-Lezama; MIT CSAIL). Surfaced via the X home feed (@omarsar0), 2026-09-26. Raw: raw/twitter/feed/2026-09-27-evening-210705-ranked.json. Background from the alphaxiv overview.

TL;DR

Agent frameworks usually bolt on subsystems: a memory store with vector search, a reflection stage, a prompt optimizer. JAZ asks how much of that is needed. It exposes one primitive, invoke. The model writes code, can call invoke recursively to spawn sub-calls, and sees all its inputs and its own execution history as variables in the code environment. Memory and self-improvement stop being separate modules and become code the agent writes at runtime. Hooks enforce constraints and monitoring. With prompting only, JAZ beats Letta (MemGPT) by 8% at half the cost on the recall-heavy part of StuLife, and beats ACE (Agentic Context Engineering, a self-improvement method) by 4% on AppWorld at lower cost.

The harness is one function the model can call

Memory and self-improvement are programs the agent writes, not modules someone installed.

flowchart LR
  I["Inputs and history<br/><small>exposed as variables</small>"] --> M["Model<br/><small>writes code</small>"]
  M --> E["Code env<br/><small>runs the program</small>"]
  E -->|invoke| M
  E --> H{"Hooks<br/><small>constraints, monitors</small>"}
  H -->|allowed| O["Result<br/><small>task output</small>"]
  H -->|blocked| X["Stop<br/><small>policy violation</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class I input
  class M,E core
  class H loop
  class O exit
  class X err
  linkStyle 2 stroke:#f08c00,stroke-width:2px

Blue is input, purple is the model and its code, amber is the hook gate and the recursive loop, green is output, red is a block.

Key points

Generalizes CodeAct and Recursive Language Models: code as the action space, plus recursive self-calls for inputs larger than one context window.
Cost falls because nothing runs unless the agent asks for it. No always-on retrieval or reflection passes.
Governance moves to hooks. The design keeps a place for enforced constraints even though the workflow is model-written.

Relation to prior wiki pages

Same week, same direction as Just-in-Time Memory (09-24). JitMem showed that summarizing a run when it ends throws away what the next task needs; it curates memory at read time instead. JAZ goes further and removes the memory subsystem entirely. See agent-memory.
Harness minimalism. agent-harness-engineering recorded on 09-22 "an argument for its own abolition." JAZ is the strongest benchmark evidence yet for that side: fewer fixed stages, better cost.

Gaps

Prompting-only results on two benchmarks; no RL-trained comparison.
Model-written memory is only as auditable as the hooks. The same week's trace-tampering paper shows why that matters.