Loop engineering represents a paradigm shift where developers design meta-systems, instrumentation, and verification gates that allow models to safely iterate and correct themselves across execution cycles without human intervention. This approach requires comprehensive logging of execution traces and verifiable performance goals, enabling agents to identify systemic weaknesses and propose meaningful modifications. Successful frameworks like Self-Harness and HarnessX implement strict regression testing and structured search pipelines to avoid "loopmaxxing"—throwing massive compute at problems without guided optimization. The developer's highest point of leverage is shifting toward designing the meta-systems and verification gates that allow models to safely iterate and correct themselves.

Researchers at Shanghai AI Lab introduced Self-Harness, a framework enabling agents to autonomously improve their operating rules through a three-stage loop: identifying failure patterns in execution traces, proposing targeted harness modifications, and validating changes through regression tests. On Terminal-Bench-2.0, lightweight models like Qwen-3.5 and GLM-5 achieved performance jumps of 33-60%. This shift moves optimization burden from developers to the AI itself, allowing smaller models to punch above their weight class while reducing API costs and inference latency. The harness—the surrounding software architecture connecting a model to tools and its environment—traditionally required manual, rigid, and time-consuming engineering; Self-Harness automates this process.

HarnessX by Xiaomi's Darwin Agent Team decomposes agent architecture into nine pluggable processors (context assembly, memory management, tool ecosystems, control flow, observability, etc.) that can be dynamically swapped without breaking surrounding code. The framework uses AEGIS, a reinforcement learning-based optimization engine, to automatically search for better structural combinations while preventing catastrophic forgetting and reward hacking. Testing on the GAIA benchmark showed a Qwen 3.5 9B model improving from 33% to 47% success rate after optimization, with code open-sourced on GitHub.

Claude Science is a dedicated research application that moves beyond discussion to actual execution of scientific pipelines, database navigation, and cluster job orchestration. The app natively connects to 60+ scientific databases including UniProt, PDB, and ChEMBL, includes a reviewer agent that flags incorrect citations and mismatched figures, and allows analyses to scale from single GPU to hundreds. A UCSF team reported certain analyses now take roughly one-tenth of the previous time. The app is available on macOS and Linux for Pro, Max, Team, and Enterprise plans.

Amazon researchers discovered a bypass of Fable 5's safety filters that allowed the model to identify software vulnerabilities and produce exploit code, triggering a US government-mandated global suspension. Anthropic restored access by implementing a new safety classifier trained on the reported bypass technique (blocking it in 99%+ of cases), temporarily routing some coding requests to Opus 4.8 while tuning filters, and collaborating with industry partners on a new jailbreak severity framework. The model is now live across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork.

Notion 3.6 brings external AI agents including Claude and Cursor into workspaces as full team members with access to databases, project boards, and documents. Agents can read and write Word, Excel, and PowerPoint files, manage Outlook inboxes and calendars, and create interactive widgets like ROI calculators visible to the entire team. Five new integrations (Mercury, Mixpanel, Miro, Box, ClickHouse) join existing ones like Figma and GitHub, with users choosing their preferred AI model from Opus 4.8 to GLM 5.2.

Claude in Chrome launched July 1st for all paid subscribers, enabling users to navigate websites, fill forms, and repeat workflows on fixed schedules without coding or API integration. The extension shares your login status across all apps where you're authenticated, from Google Sheets to CRM systems. Pro subscribers (€20/month) run only Haiku 4.5; complex tasks require the Max tier.

Nemotron-Labs-TwoTower splits a single 30B model into two copies: one processes the prompt and maintains context, while the other fills masked text slots in parallel and refines predictions over multiple passes before committing. The technique achieves 2.42x faster generation versus the original model while retaining 98.7% quality on benchmarks like MMLU, GSM8K, and HumanEval, running on 2x H100 or A100 GPUs (59GB per GPU). The model is available via Hugging Face Transformers and requires no training from scratch.

This development challenges the prevailing assumption that larger models always perform better, demonstrating that efficiency and architecture design can overcome raw parameter count on specific task categories like long-horizon agent planning. The result signals a shift in the industry away from pure scale toward smarter model design and routing strategies.

GitHub Copilot expanded model choices to include open-weight Kimi K2.7 Code, offering a more cost-effective option for coding tasks. This continues the trend of multi-model routing, allowing developers to select the cheapest capable model for their specific task.

GitHub Copilot CLI and SDK now support configurable AI credit limits per session, preventing unattended automations from silently consuming budgets. This reflects the broader industry shift toward cost control and budget visibility as AI tool spending accelerates.

Microsoft raised M365 pricing effective July 1 across most tiers: Business Basic increases 16% to $7/user/month, Enterprise E3 rises 8% to $39/month. Business Standard and Premium bundles now include Copilot as standard offerings ($23.50 and $32 respectively), eliminating the need for separate Copilot licenses. Standalone Copilot Business licenses increase from promotional $18 to $21/user/month, while existing customers retain current pricing until contract renewal. This bundling strategy signals Microsoft's confidence in Copilot adoption and its shift from optional add-on to core product.

A coalition of 30 MEPs from six political groups, led by Dutch parliamentarian Reinier van Lanschot (Volt), is calling on the European Council to organize a special AI summit by October 15. The initiative highlights Europe's strategic dependence on foreign AI providers, citing recent U.S. export restrictions that made Fable 5 unavailable for weeks as evidence. Van Lanschot advocates for a coordinated European approach centered on public values, noting Europe has significantly less computing power than the U.S., with support from former Estonian president Toomas Hendrik Ilves.