MIT paper: a minimalist agent loop that passes history as code variables beats Letta and ACE at half the cost

rohanpaul_ai · x · 2026-10-07

A new MIT paper, "Harness as a Language," introduces JAZ: an agent that treats its own prompt and full history as code variables it can search and pass by reference to subagents, instead of manually copying context.
Typical setups bolt on long-term memory (e.g., Letta) or self-improvement systems (e.g., ACE); these add cost and break when tasks don't fit their design.
JAZ keeps only the loop: the model writes Python, calls itself as a subagent, and passes prompt and history by reference so no information is lost.
Results: on StuLife tasks requiring facts from 50+ earlier tasks, JAZ hit 69.9% vs Letta's 61.8% at less than half the cost; across 417 AppWorld tasks, self-improvement reached 74.2% vs ACE's 69.9%, again for less.
Takeaway: a maximally expressive minimalist harness can outperform purpose-built memory and self-tuning systems at lower cost.

More from coding & agent