Weekly collection of AI engineering news that I found interesting. Filtered from a pile of search results, narrowed to what actually matters if you build things.

Nine models in four weeks

February and March 2026 will probably get a Wikipedia section. Claude Opus 4.6 dropped February 5. Then GPT-5.4 on March 5. Between those two, Google shipped Gemini 3.1, Meta dropped Llama 4 Maverick with a 10M token context window, and several Chinese labs pushed major open-source releases. One YouTube channel called it “the most explosive month in AI history.” That bar seems low given the past two years, but the pace does feel different now.

GPT-5.4 is the one worth looking at this week. OpenAI describes it as their most capable model for professional work. It combines reasoning with coding and what they call “agentic workflows” into a single model. A few things stand out. Context window hits 1M tokens. Hallucinations are reportedly down 33% compared to GPT-5.2. And it’s the first OpenAI model with native computer use: it can click, type, and navigate software by interpreting screenshots. On OSWorld, a benchmark for desktop task completion, it scored 75%, slightly above the human baseline of 72.4%.

One practical change for API users: OpenAI introduced Tool Search, a system that looks up tool definitions on demand instead of stuffing the entire tool list into every prompt. If you’ve ever debugged a bloated system prompt because you have 40 tools defined, you’ll understand why this matters.

Claude Opus 4.6 (which I’m running right now) still leads SWE-bench coding benchmarks and added 1M token context in beta along with agent team coordination. The gap between models has narrowed so much that benchmark wars are starting to feel like fuel economy claims on car ads.

The benchmark war is getting silly

Every release now comes with a table of numbers proving superiority on at least five metrics. Some of those metrics exist primarily to win those tables.

The Plandek Engineering Productivity Benchmarks, released March 5, offer something more useful: real data from 2,000+ engineering teams. The headline finding is counterintuitive. AI tools help low-performing teams 4x more than high-performing ones. Low-performing teams cut lead time by nearly 50% using AI. High-performing teams see a 10-12% improvement.

That’s not a knock on AI tools. It tells you something about where the ceiling actually sits. High-performing teams already have good processes and fast feedback loops. AI fills gaps. Low-performing teams have more gaps to fill.

A separate study from The AI Corner found that experienced developers took 19% longer on tasks when using AI coding tools, despite believing they were 20% faster. The perception gap is real. Whether that comes from overconfidence in AI output quality, or from the overhead of reviewing AI-generated code carefully, isn’t clear. Probably both.

My take: AI coding tools are not shortcuts. They’re leverage, which is different. Leverage amplifies what you already do well, and also what you do poorly.

Enterprise is buying in, governance is not keeping up

Deloitte published their State of AI in the Enterprise report for 2026. Roughly 90% of enterprises now use AI in some form. Only about 20% have redesigned core workflows to incorporate it. The rest bolted it on.

The governance gap is the bigger story. Shadow AI is the new shadow IT. Marketing teams are connecting generative AI to CRM exports. Developers call third-party model endpoints without data classification controls. Security teams find out later. Gartner predicted that over 80% of enterprises would use generative AI APIs in production by 2026. The governance frameworks, they noted, lag adoption.

This is not a new pattern. It’s how cloud adoption went in 2012-2016. Every new capability gets deployed first, governed second. The difference with AI is the risk surface. A misconfigured S3 bucket leaks data. A misconfigured AI agent with access to your production systems can make decisions you can’t easily audit or reverse.

The practical answer from the enterprise side: identify high-risk AI applications first, build governance into deployment rather than retrofitting it later. Straightforward advice that most organizations still aren’t following.

Agentic AI: reality check

Multi-agent systems are the trend everyone is writing about. The prediction, as the Moxo agentic trends report puts it, is that 2025 was the year of the AI agent and 2026 is the year of multi-agent systems. Procurement workflows with a squad of specialized agents. Legal review agents handing off to compliance agents. A manager agent coordinating the whole thing.

I run multi-agent setups myself with Claude Code agents and skills, so this isn’t theoretical. The coordinator pattern does work. The coordination overhead is real too. Shared task lists, message passing, context handoffs between agents. When it breaks, debugging is harder than debugging a single agent.

The apmdigest piece from Xurrent is worth reading for the cold water: “In 2026, there’s going to be a significant gap between what vendors market and what IT leaders actually allow AI agents to do.” The blast radius of a misconfigured autonomous agent on production infrastructure is not like the blast radius of a wrong config file. IT teams are right to be cautious.

My read: multi-agent works well for narrow, well-defined workflows with clear handoffs. It works poorly when the task scope is vague or the state management between agents gets complex. That’s not a criticism, it’s just where the technology is today.