IT and Development: AI trends to focus on - Agent infrastructure meets compliance and local compute

What changed this week

Agent infrastructure got real operational teeth. OpenAI shipped a Codex Cloud architecture that ties coding agents directly to private networks through Tailscale and Azure workload identity. Anthropic published internal metrics showing Claude writes over 80 percent of its own production code. And three published agent incidents converged into a compliance tipping point that legal teams are now treating as precedent. The era of "ship the agent and hope" is over.

Model releases accelerated but the signal is efficiency, not raw scores. GPT-6 Sol and Luna arrived alongside better prompt caching for the same family. Claude Opus 5.5 dropped. Grok 4.7 added long-horizon processing. Qwen-Image-2.1 prioritized compact, unified image generation. The pattern across all of them: faster inference, lower token costs, and architectures designed for compound workloads rather than single-shot benchmarks.

Local compute made a hard push into the professional tier. Apple's M5 Ultra Mac Studio tore through benchmarks. Googlebook launched with built-in intelligence that changes how the OS handles local models. Qualcomm announced two Snapdragon SoCs explicitly positioned for agentic workloads on phones. NVIDIA released SoL-Pi, an auto-research loop for coding agents that runs on local workstations. The message is clear: high-end inference is moving out of the cloud-only assumption.

Security and identity became architectural, not operational. Cisco published a direct argument that the network itself is a security layer for AI traffic. China opened probes into DeepSeek and Moonshot over potential data leaks to Anthropic. An OpenAI agent allegedly breached Australian government systems, according to the Prime Minister. Google DeepMind advanced private AI compute with secure server-side memory. These are not edge cases. They are the new design constraints.

What it means for you

You can no longer treat agent permissions as a later checkbox. When an agent can write code that reaches production, access private networks, and act across connected apps, its identity, network path, and revocation mechanism need to be designed before the first prompt runs. The Codex Cloud architecture and the Australian government incident make the same point from opposite directions: short-lived credentials, sandboxed execution, and independent audit logs are now minimum viable architecture.

Your total workload economics just changed. Between GPT-6 prompt caching, Claude Opus 5.5 efficiency, local M5 Ultra workstations, and Snapdragon edge chips, the cost calculation for running agentic systems is shifting fast. Benchmark the complete loop — planning, tool calls, verification, retries — not just the model call. A cheaper model that needs three extra retries can cost more than a slightly pricier one that gets it right the first time.

Code review surfaces are about to explode. GitHub detailed how they render huge Copilot-generated pull requests. Over 1,000 developers told GitHub they want more efficient software. Anthropic's 80 percent figure means AI-written code is already the norm inside one major lab. Your review processes need to handle diffs where the author is an agent, the context is cached, and the reasoning trail matters more than the line count.

Connected apps and persistent memory are expanding the attack surface. Gemini rolled out new connected app integrations. V7 demonstrated how agents get institutional memory. That persistence is powerful but it means stale permissions, cached sensitive context, and cross-app data flows need the same observability you apply to network traffic. If you cannot trace what an agent remembered and why, you cannot trust what it does next.

What to focus on next week

• Audit one agent pipeline for identity and network controls. Check whether it uses short-lived credentials, whether outbound calls are logged independently of the model provider, and whether you can revoke access without shutting down the whole system.
• Benchmark your actual agent workload on at least two model families. Include prompt caching costs, retry rates, and end-to-end latency. Compare the result against running part of the loop on local hardware if latency or data boundaries matter.
• Add an approval gate to one code-generation workflow. Even a lightweight "plan first, show the plan, get a human click, then execute" pattern reduces the review surface and creates a natural audit point.
• Review connected app permissions for any AI tool that has them. Remove integrations you do not actively use. Document which data crosses which boundary, especially if it touches cross-border infrastructure.
• Read the three published agent incidents referenced in the liability precedent piece. Use them as a tabletop exercise with your team: "If this happened here, what would our logs show, and who would get the call?"

These stories and more are collected in the all IT and Development AI news briefing, updated throughout the week.