Weekly Dose #11 - AI Agents Are Getting Easier to Build, and Harder to Control

2025-01-29 · Source: Machine Learning Pills · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cybersecurity & Data Privacy · Depth: Intermediate, long

Summary

The Weekly Dose #11, covering July 3 to July 18, 2026, highlights significant advancements and security concerns in AI/ML. OpenAI released its GPT-5.6 family (Sol, Terra, Luna) on July 9, integrating programmatic tool calling and multi-agent orchestration, with Sol priced at \$5 per million input tokens and \$30 per million output tokens. PyTorch 2.13, released July 8, introduced "nn.LinearCrossEntropyLoss" for up to four times GPU memory reduction and FlexAttention on Apple Silicon for up to 12 times speedup. Meta launched Muse Spark 1.1 and its Model API on July 9, while SpaceXAI released Grok 4.5 API on July 8 and open-sourced Grok Build on July 15, intensifying competition in the agentic work layer. Google enhanced Gemini API's Managed Agents on July 7 with asynchronous execution and persistent state. Concurrently, Wiz disclosed GhostApproval on July 8, a critical vulnerability affecting several AI coding assistants like Amazon Q Developer and Claude Code, where symbolic links could bypass human approval for sensitive file writes.

Key takeaway

For AI Engineers evaluating new agentic systems, you must broaden your assessment beyond base model capabilities. Focus on benchmarking complete workflows, including orchestration, tool routing, and caching, to understand true costs and performance. Ensure your security protocols for agents explicitly verify canonical paths and log all accessed resources, as human approval is insufficient without accurate provenance. Prioritize framework updates like PyTorch 2.13 for significant efficiency gains in training and inference.

Key insights

AI platforms are increasingly integrating orchestration and execution, shifting competition beyond base models.

Principles

• Orchestration is becoming a platform primitive.
• Agent harnesses are competitive layers.
• Human approval requires accurate provenance.

In practice

• Agent performance depends on model, harness, context, and tools.
• Fused operations and communication overlap improve model economics.
• Long-running agents require resumability and persistent state.

Topics

• AI Agents
• Model Orchestration
• PyTorch Optimizations
• Agent Security
• Large Language Models
• Distributed ML

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Weekly Dose #5 - AI Is Building AI, and Your Support Bot Just Reset a Password — Agents are rapidly transitioning into authority layers, with that authority increasingly moving to the edge.
• [AINews] not much happened today — Frontier model constraints drive multi-model orchestration and advanced agent infrastructure for robust AI applications.
• LAI #116: Agents Are Easy. Operating Them Isn’t. — Operating AI agents reliably in production requires addressing GPU economics, scalable function calling, and secure execution.
• This Week in AI: Fable 5, the Clone Wave, and Uber’s AI Reality Check — Efficient AI agent development hinges on leveraging existing "ingredients" and building agent-native, observable tooling to manage costs.
• not much happened today — AI development is converging on durable, inspectable agent execution, pretraining efficiency, and enterprise-grade security.
• [AINews] Open Models, Model Labs vs Agent Labs, and What's Untrainable — Sarah Guo — The AI landscape is rapidly evolving with open models, agentic capabilities, and new architectures, but trust and "intent" remain critical challenges.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning Pills.