More Model Choices, More Work Around the Code

With an RBC survey finding more than half of respondents running AI workloads in production, model selection, agent controls and the work surrounding code generation become concrete engineering concerns.

Model selection gains more dimensions

OpenAI's GPT-5.6 Preview introduces Sol, Terra and Luna, apparently aimed at different capability and cost tradeoffs, with safety documentation describing staged access. Model choice may therefore depend on availability as well as performance. Access restrictions are also shaping alternatives: Asian startups are launching Mythos-like models while Anthropic's export ban continues, including Sakana AI and 360 Security. Semgrep reports that GLM-5.2 beats Claude in its cybersecurity benchmarks, giving security teams another candidate to evaluate for those tasks.

Deployment constraints add another consideration. Google Research describes frozen multi-token prediction speeding Gemini Nano on Pixel hardware, relevant to latency-sensitive and offline features. In media tooling, Adobe is acquiring Topaz Labs, known for AI image and video enhancement; terms were not disclosed. The acquisition would bring that specialized technology into Adobe's ownership, with possible implications for Creative Cloud workflows.

Production agents need operational control

Understanding production use requires more than adoption counts. Anthropic's June Economic Index analyzes aggregated Claude usage to describe how businesses and individuals use its models, offering a reference for adoption patterns within Claude's user base. On the operational side, Okta is extending AI agent governance to FedRAMP and HIPAA customers, bringing identity, access and audit controls to agents in government and healthcare environments.

Identity controls address only part of agent operations. Simon Willison's account of roughly 2,000 people attempting to hack his assistant documents prompt-injection and jailbreak attempts against a working system. An essay on who can actually shut down deployed AI argues that stopping a system can span API access, infrastructure dependencies and contractual obligations. These raise concrete design concerns: how an agent behaves under adversarial input, and who has the authority and means to halt it.

Repository context becomes agent infrastructure

Coding agents need project state as well as source files. GitLab Orbit connects Google's Antigravity agents to issues, merge requests and CI state, expanding the context available for implementation decisions. Version control itself also has a new option: Epic Games released Lore as an open-source version control project. Orbit addresses the information available to an agent; Lore introduces another version-control project for teams to evaluate.

More output puts pressure on judgment

VentureBeat reports that Claude Code-style tools are turning individual engineers into three-person-equivalent output units, while exposing a shortage of product thinking. If implementation capacity rises, deciding what deserves implementation may become a tighter constraint. An organizational argument proposes smaller, more capable teams rather than headcount cuts across existing structures, emphasizing focused ownership as tooling changes throughput.

A practitioner's account of engineering with AI brings that discussion back to planning, code review and debugging. Those workflow changes give the staffing debate a concrete foundation: generating code is one part of delivering software, alongside decisions about product direction, review and diagnosis. Smaller teams remain a proposal for organizing that work.