What Enterprise "Agents" Really Are: 71% Are Chatbot Wrappers

VentureBeat surveyed 101 enterprises (100+ employees) on the reality of agent orchestration, and the results surface some striking numbers.

**Anthropic Claude dominates platform share**

In agent orchestration, Anthropic's Claude leads with 40%, far ahead of second-place Microsoft (18%) and third-place OpenAI (13%). The biggest factor in platform selection is "model gravity" — that is, foundation model performance — while "task completion reliability" (32%) and "multi-step workflow management" (28%) top the list of success metrics.

**But the reality is harsh**

The most striking finding: 71% of enterprises reported that a quarter or fewer of their deployed agents perform true multi-step orchestration. Only 10% said more than half do. Most of what gets called an "agent" is nothing more than a single-prompt chatbot wrapper.

On top of that, 27% of enterprises have no cost controls and admit they have no way to stop a runaway agent before the invoice arrives. Vendor lock-in (35%) is the top concern, and 51% expect a hybrid control plane by year-end.

Infrastructure and budgets for the orchestration layer are coming together, but the maturity of the agents it's all meant to orchestrate is lagging behind — that seems to be the current state of enterprise AI.

MIT Paper Warns of "Speculative Growth" in AI Investment

A paper from MIT's economics department, "Speculative Growth and the AI 'Bubble'", has been published and is drawing discussion on Hacker News with 33 points.

The paper analyzes the AI sector's investment rush as "speculative growth," arguing that the current AI investment boom may exhibit the characteristics of a speculation-driven bubble rather than sustainable growth. The BIS (Bank for International Settlements) has also released a working paper titled "The AI Investment Race," adding to the growing analysis of the macroeconomic impact of the AI investment race.

These papers offer a cooler, economically grounded perspective amid the feverish phase of AI investment.

"We Don't Use AI" — A Design Company's Clear Stance

Mass Driver, a design and manufacturing company, published an article titled "We don't use AI in any of our design or production processes," earning 49 points and 25 comments on Hacker News.

Companies that openly declare they don't use AI are still a minority, but the heat of the discussion shows that a critical stance toward AI adoption is winning real support within the tech community. It's a fascinating signal that we've shifted from "of course we use AI" to "should we be using AI at all?"

A Mid-Training Method for Coding Agent Foundation Models

"Function-Aware Fill-in-the-Middle (FIM) as Mid-Training for Coding Agent Foundation Models" was featured on Hugging Face Daily Papers.

The paper shows that FIM mid-training on functions selected through program dependency graph analysis consistently improves SWE-Bench scores. What's particularly notable is the finding that the structure of ordinary code — "function call → return value → downstream use" — is structurally isomorphic to a coding agent's "action → observation → continuation" steps. Based on this, the authors argue that source code itself can serve as a training signal for agents, without relying on scaling synthetic agent trajectories.

The authors have released everything: a corpus of 400K samples (2.6 billion tokens) extracted from 968 repositories, the pipeline, and the checkpoints.

Local LLM Community: A New Era of "Mini-Boss" Models Running on Dual DGX Sparks

Discussion is heating up in Reddit's LocalLLaMA community about running large models at 4-bit quantization on two DGX Sparks (~$8,000), which provide 250GB of usable memory.

Until recently, the recognized models in this size class were limited to the likes of GLM 4.5/4.6/4.7 (194GB) and Qwen 3.5 397B (210GB), but the past three months have brought:

MiniMax M2.7** (131GB)

DeepSeek V4 Flash** (160GB) — reported to deliver strong performance even at 2-bit quantization

Xiaomi MiMo 2.5** (181GB)

StepFun 3.7 Flash** (129GB)

Tencent Hy3** (182GB)

$8,000 is no small sum, but the fact that the environment for running models of this class locally is falling into place is a sign of the open model ecosystem's maturation.

---