AI Agent Enters the Harness Engineering Era: From Model Competition to Infrastructure Restructuring

In 2026, the AI Agent field is undergoing a paradigm shift from model capability competition to infrastructure engineering. Research and practice from multiple institutions indicate that the key bottleneck for Agent performance has shifted from the model itself to the outer system wrapping the model—the Harness (constraint engineering/control shell).

Harness Engineering: A New Lever for Agent Performance

The Self-Harness framework proposed by Shanghai AI Laboratory enables models to automatically mine failure modes based on their own execution trajectories, propose Harness modifications, and validate adoption through regression testing. On Terminal-Bench-2.0, modifying only the Harness without changing the model resulted in a total improvement of 104% for Qwen3.5-35B-A3B, and 28% and 24% for MiniMax M2.5 and GLM-5, respectively. This work was retweeted by the CEO of LangChain and included by former OpenAI VP Lilian Weng.

Another framework, [schema] from Impossible Research, UC Berkeley, and CMU, achieved 98.98% RHAE on the ARC-AGI-3 Public set when combined with Claude Opus 4.8, and 95.35% with GPT-5.6 Sol. Its core idea is to write the world model as runnable, verifiable programs, allowing LLMs to reason through hypothesis-experiment cycles like physicists. ARC Prize President Greg Kamradt acknowledged the approach but noted that some scores rely on fixed fallback rules and have not yet been independently verified by the official team.

Semantic Layer and Data Foundation: Key Support for Agent Deployment

A Data for AI community event in Beijing pointed out that without a semantic layer, Data Agent accuracy is stuck at 70%; with a proper semantic layer, it can approach the 95% production-ready line. Datastrato founder Du Junping proposed that data architecture is shifting from engine-centric to metadata-centric, with metadata systems evolving from passive query books to active control planes.

Apache Ossie (formerly Open Semantic Interchange) entered the Apache Incubator on July 8, aiming to standardize semantic metadata exchange formats and provide a unified business language for AI Agents. Its ai_context field can supplement Agents with instructions, synonyms, and examples, helping map natural language to database names.

Computing Power and Architecture Innovation: Supporting Agent Scale

At WAIC, SenseTime released a heterogeneous hybrid inference solution that splits inference into Prefill (reading the question) and Decode (answering), using domestic GPUs for Prefill and high-end GPUs for Decode, making hybrid inference clusters with domestic chips profitable. Its daily token output was 400 billion at the beginning of the year, expected to reach 2.42 trillion by the end of July, with a year-end target of 10 trillion. It also launched SenseNova U1 Pro, positioned as a delivery-grade native multimodal Agent foundation for long-horizon tasks, initially targeting the design industry with a delivery rate exceeding 60%.

Tencent Cloud's Big Data AI DLC launched the first production-grade fully managed Spark+Ray platform in China, allowing data lakes to simultaneously support two engines, enabling Agent preprocessing, fine-tuning, and inference to be closed-loop on a single platform.

Skills and Memory: New Dimensions of Agent Capability

The SkillsBench benchmark dataset (84 tasks, 11 domains) quantitatively evaluates the real effectiveness of skills (SKILL.md): manually curated skills improve task pass rates by an average of 16.2 percentage points, while AI self-generated skills yield no positive gains. Skills that are concise and single-responsibility provide the highest gains, while stacking multiple skills tends to introduce context noise.

In memory systems, the MemCon framework models memory management as a control problem, using UCB bandit to adaptively select retrieval strategies, improving success rates by up to 15.2 percentage points across six tasks. The A-TMA framework reveals the ghost memory problem—where outdated facts coexist with current facts, causing temporal confusion.

Commercial Deployment: From Tools to Digital Employees

Institutions like ModelBest open-sourced the StaffDeck digital employee platform, treating AI Agents as formal employees with employee IDs, positions, SOPs, and performance data. It uses state-machine-driven process skills, OKF-structured knowledge bases, and feedback iteration loops to address issues like process confusion and ungrounded responses in real business scenarios.

Tec-Do released Navos 2.0, upgrading its multi-agent product to an agentic workflow architecture covering the entire chain from product selection, creative design, to influencer outreach. One person plus AI can accomplish the work of a traditional team. Nile targets Agent e-commerce infrastructure, helping merchants build AI-native product systems with conversion rates reaching 150%-300% of in-store rates.

Recursive Self-Improvement: Self-Evolution of Harness

In her latest survey, Lilian Weng systematically outlined the path to recursive self-improvement (RSI) driven by Harness Engineering: through paradigms such as automated workflows, file-system-persistent memory, and sub-agent parallelism, models can autonomously optimize their outer constraint systems. This has been validated on code and research benchmarks, but still faces seven major bottlenecks, including difficulty in evaluation quantification, insufficient long-term memory, and reward hacking.