MIT CSAIL introduces JAZ minimalist agent framework

// 2h agoRESEARCH PAPER

MIT CSAIL introduces JAZ minimalist agent framework

Researchers from MIT CSAIL have introduced JAZ, an open-source agent framework that replaces heavyweight orchestration graphs and external memory databases with a bare Python REPL and a single recursive invoke primitive. Described in the paper "Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity," JAZ exposes all model inputs, tools, and execution history directly as in-memory Python variables within the runtime environment. By empowering the language model to write arbitrary code and recursively call invoke on subproblems, the framework unifies context management, tool dispatch, and agentic workflows without bespoke scaffolding. Across empirical evaluations, JAZ outmatches specialized agent systems like Letta on long-horizon recall and ACE on continuous self-improvement workflows while substantially reducing token overhead and operational cost.

// ANALYSIS

The agent ecosystem's obsession with heavyweight graph orchestrators and external memory databases was premature optimization; treating the harness as an interactive Python REPL with a single recursive invoke primitive is fundamentally the right abstraction.

–Collapses multi-layer orchestration into code: Instead of rigid JSON schemas, prompt chains, or bespoke state machines, JAZ relies on native Python syntax and recursion to express complex agentic topologies.
–Memory as standard state: Exposing interaction history directly as runtime variables eliminates vector database retrieval latency and semantic chunking failures on long-horizon recall.
–Programmatic recursion: Implementing the agent loop as a first-class language primitive allows models to spawn sub-agents and iteratively refine tasks without bespoke orchestration wrappers.
–Significant cost and token savings: Slashing architectural bloat and boilerplate system prompts yields higher benchmark accuracy while substantially reducing prompt token consumption compared to specialized frameworks.
–Sandboxing challenges in production: While executing arbitrary Python code maximizes expressivity, deploying JAZ in untrusted environments requires strict sandboxing and execution guardrails.

// TAGS

jazagentopen-sourceframeworkagent-memorydevtoolresearch

DISCOVERED

2h ago

2026-09-25

PUBLISHED

2h ago

2026-09-25

RELEVANCE

9/ 10

AUTHOR

Discover AI

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

POLICY1h ago

NSA spends billions testing frontier AI models

According to sources familiar with classified intelligence estimates, the National Security Agency informed lawmakers that it is spending billions of dollars in taxpayer funds this year to evaluate and stress-test advanced artificial intelligence models for national security vulnerabilities. The classified price tag—funded through defense intelligence budgets and far outstripping prior congressional estimates of tens of millions for federal AI oversight—indicates that comprehensive government evaluation of cutting-edge models could cost tens of billions annually.

INFRA1h ago

Merge Gateway cuts GLM 5.3 pricing

Merge has announced the final week of a 50% promotional discount on GLM 5.3 through Merge Gateway, with the special rate set to expire on September 30. Developed by Zhipu AI, GLM 5.3 scores 60 out of 100 on the Artificial Analysis index and delivers faster inference speeds than other models in its performance class.

NEWS2h ago

First principles thinking unlocks agentic AI engineering

In this essay, Sunil Sadasivan reflects on the "senior engineer death spiral" and argues that overcoming stagnation requires first principles thinking and a relentless focus on momentum over outcomes. He explains that senior engineers frequently become constrained by accumulated experience and outdated technical boundaries, which can hinder adaptation during the transition to AI agent-driven development. By temporarily setting aside historical assumptions—"putting experience in a box"—engineers can view problems with fresh eyes, leverage AI agents for rapid feedback loops, and achieve a new flow state rooted in fundamental problem understanding.

MIT CSAIL introduces JAZ minimalist agent framework