U
Unnati Tripathi
@aiwithunnati

Master Harness Engineering

3 must-read papers & articles on AI agent harnesses, context engineering, and long-horizon reliability curated for builders.

1
Anthropic Engineering Blog
Effective Harnesses for Long-Running Agents

Normal agents work fine at 200 lines of code but push them to 2000 lines and they completely lose track of what they were doing. Anthropic took inspiration from human engineers working in shifts and designed a better two-part harness to fix exactly this.

The key insight: agents need to leave structured artifacts (git commits, progress logs, feature lists) so the next session knows exactly where to pick up.
Read Article
2
Research Paper · NKU · April 2026
Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents

You keep adding memory to an agent thinking it will perform better. But this paper proves that by the 8th9th step, performance starts crashing badly. Capability and reliability are two very different things and most benchmarks only measure capability.

The paper introduces 4 new metrics: Reliability Decay Curve, Variance Amplification Factor, Graceful Degradation Score, and Meltdown Onset Point.
Read Paper
3
Anthropic Engineering Blog
Effective Context Engineering for AI Agents

Prompt engineering is no longer just about finding the right words. Context engineering means optimally curating all the tokens available to the LLM so it consistently produces the desired behavior across every step of a long task.

Context is a critical but finite resource. The best agents aren't powered by the best model they're powered by the best context management.
Read Article