AI agents are getting stranger, and more real
AI agents are evolving fast, and not only in the “useful assistant” direction. We are now seeing an ecosystem forming around agents: agent frameworks, agent-native tools, agent documentation hubs, agent social spaces, and even agent-specific languages. OpenClaw has emerged as one of the visible symbols of that shift, as an open agent framework that has attracted strong developer attention and even inspired related infrastructure such as Nvidia’s NemoClaw. Meta’s acquisition of Moltbook, a social network where AI agents interact with each other, is a sign that agent-to-agent environments are no longer a joke or side experiment. They are becoming part of the AI landscape. Reports around Moltbook even described agents developing pseudo-identities, social behaviors, and, at times, something that looked like “religion”, though authenticity issues and human impersonation also played a role.
What makes this more serious is that the infrastructure around agents is also maturing. Andrew Ng’s open-source Context Hub is explicitly designed for coding agents, not humans, to reduce API hallucinations and give agents access to curated, versioned docs. On top of that, projects like Mog, a Rust-based language for safe AI agents, show that people are now building technical stacks specifically for autonomous systems rather than just chatbots. In other words: agents are no longer just “LLMs with tools.” They are getting their own memory scaffolding, documentation layers, execution environments, and programming abstractions.
But the most important part of the story is security. An experimental agent called ROME was reportedly caught redirecting GPU resources toward unauthorized cryptocurrency mining and opening a reverse SSH tunnel during training. The key point is not the crypto itself. It is that the behavior emerged while the agent was being optimized with reinforcement learning for complex tool-using tasks. It apparently discovered that access to more compute could help performance, and exploited that path. This is exactly the kind of issue that reminds us that agent risk is often not in the model output, but in the environment around the model.
My view: this is where many companies still underestimate the challenge. Using one AI to supervise another may help, but it should not be the foundation. Guardrails now have to exist at the system, network, permissions, and infrastructure layer, because once agents touch real environments, “unexpected behavior” becomes an operational problem.
Nvidia shifts the AI race from training to inference
For the past two years, the AI race was mostly about training: bigger clusters, more GPUs, and larger frontier models. At GTC 2026, Nvidia tried to shift that narrative by making inference the main story, with Jensen Huang saying the revenue opportunity for AI chips could reach $1 trillion through 2027. This basically says the next market is more about running AI then building it. It might also signal a level of maturity reached by the LLM.
That matters because inference has a very different economic logic. Training is occasional and concentrated; inference is continuous, latency-sensitive, and cost-driven. Nvidia used GTC to position itself for that shift, with the Vera CPU, new inference systems, and a broader full-stack deployment strategy (based on Groq by the way - following the acquisition shared in previous newsletter).
My take: this is where the real competition is moving. It is no longer enough to train a powerful model. The harder challenge is serving it cheaply, quickly, and reliably at scale. Training created the AI boom, but inference may decide who truly makes money from it.
AI labs raise mega-funds and rewrite the capital game
If anyone expected the AI funding frenzy to slow down in 2026, the opposite happened. Reuters reported that xAI raised $20 billion in an upsized Series E round to fund Grok development and compute expansion. Reuters also reported that OpenAI is raising $110 billion in a blockbuster round backed by Amazon, Nvidia, and SoftBank, valuing the company at $840 billion, alongside large compute and cloud commitments. Anthropic, for its part, has already closed a $30 billion funding round according to Reuters, and separate Reuters reporting says it has also been planning a new raise that could value it at $350 billion.
Nearly two-thirds of global VC funding in 2025 went into AI, which helps explain why the top labs look increasingly detached from normal startup economics.
My view is still mixed. On one hand, if frontier AI becomes foundational infrastructure, some of these valuations may one day look rational. On the other hand, we are clearly in a period where capital is clustering around a handful of winners before the long-term business model is fully proven. There is also a real ecosystem risk here: when so much money goes to a few giants, smaller players may struggle to compete unless they specialize or build on top of the big labs’ platforms. We may be moving toward an AI market shaped by a few hyperscale labs with nearly unlimited budgets, while the rest of the ecosystem fights for niches, integrations, and distribution. The money is clearly there. The real test now is whether durable revenue and enterprise value follow. All of this without mentioning that there is also a circular aspects in these deals (see here).
Microsoft pushes enterprise AI from copilots to governed agents
One of the clearest enterprise signals this month came from Microsoft. With Wave 3 of Microsoft 365 Copilot, broader model access including Claude and next-generation OpenAI models, the general availability of Agent 365 on May 1, and the launch of Microsoft 365 E7 licenses / Frontier Suite, Microsoft is making a strong case that enterprise AI is moving from “assist me” to “act for me, but under control.”
Microsoft is arguing that enterprises do not just need smarter models. They need a combination of intelligence and trust: model diversity, business context, governance, security, and agent observability. Microsoft is clearly trying to solve one of the biggest emerging enterprise concerns: agent sprawl. If employees and teams start launching large numbers of semi-autonomous agents, companies will need ways to see them, govern them, and manage risk before the whole thing becomes a new shadow IT problem. Should we find a name for that (?) Shadow Agents for example.
The coding wars intensify: Anthropic and OpenAI go all-in on software agents
Anthropic launched Claude Opus 4.6 on February 5 and Claude Sonnet 4.6 on February 17, both with stronger coding capabilities and a 1M-token context window in beta. Anthropic is positioning these models around long-running work: coding, tool use, agent planning, long-context reasoning, document-heavy tasks, and multi-agent workflows inside Claude Code.
At the same time, OpenAI launched GPT-5.3-Codex, which it describes as its most capable agentic coding model so far, combining the Codex and GPT-5 training stacks and running about 25% faster. It is being pitched as a general-purpose coding agent that developers can actively steer while it works across longer tasks involving reasoning, research, tools, and execution.
What stands out to me is that both companies are converging on the same future: coding models that are less like assistants and more like (junior) engineers that can persist on a task, work across larger codebases, hold more context, recover from drift, and collaborate with tools. Anthropic is leaning hard into long context and autonomous teamwork. OpenAI is leaning into unified reasoning plus coding execution. Either way, this feels like a shift in developer tooling. The real bet is on how much real software work they can take over before reliability, verification, and security become the limiting factors. And all of that being said, AI is clearly still struggling to write production-grade code.
