Agents, Infra, and the New Speed Race: Today’s AI & Data Rundown (Feb 17, 2026)

Today’s theme is practical AI: faster inference for better UX, safer agent integration patterns, and the infrastructure choices that make systems predictable in production.

TL;DR

• DataCamp unlocked its full curriculum for free from Feb 16–22, 2026 (no credit card required).
• A local “AI hub” stack (Docker + Ollama + n8n) highlights the growing local-first automation trend.
• Low-latency inference is becoming a differentiator, with multiple providers competing on speed and throughput.
• Feature stores are resurfacing as real-time ML and agentic systems raise the stakes on freshness and consistency.
• Two “agent plumbing” stories stood out: human-in-the-loop interrupts for safer execution, and Agoda’s open-sourced approach to turning APIs into MCP servers.

DataCamp “Free Access Week” (Feb 16–22, 2026): full curriculum unlocked

What happened
DataCamp is running a “Free Access Week” offering 100% free access to its entire curriculum from Feb 16–22, 2026. The promotion highlights interactive courses, projects, career tracks, and certification prep, and states no credit card is required.

Why it matters
Time-boxed access changes the decision for would-be learners: it’s a low-friction way to explore Python, SQL, and BI tooling without committing to a subscription. It also underscores how competitive data/AI education has become as “AI-adjacent” literacy moves toward a baseline skill for analysts.

Key details

• Free Access Week runs Feb 16–22, 2026.
• The post claims the entire curriculum is free during the window.
• The post states no credit card is required.
• Promotion emphasizes courses, career tracks, projects, and certifications.

Source links
https://www.kdnuggets.com/datacamp/02/2026/learn-python-sql-and-powerbi-to-become-a-certified-data-analyst-for-free-this-week

Self-hosted AI hub with Docker + Ollama + n8n (private/local automation)

What happened
A KDnuggets guide outlines a beginner roadmap to build a self-hosted “AI hub” using Docker (and Portainer), Ollama for running local models, n8n for workflow automation, and Nginx Proxy Manager for access and routing. The guide positions the stack as a privacy- and control-oriented alternative to cloud-only workflows.

Why it matters
Local-first is becoming a real product choice: predictable costs, fewer data-handling concerns, and the ability to automate against local files and services. The flip side is operational responsibility—hardware, updates, security, and uptime move onto the builder.

Key details

• Proposed stack includes Docker/Portainer, Ollama, n8n, and Nginx Proxy Manager.
• The guide frames benefits as “no cloud fees,” more privacy, and more control.
• The article suggests modest x86-64 hardware can work, citing 8GB+ RAM as guidance.

Source links
https://www.kdnuggets.com/self-hosted-ai-complete-roadmap-for-beginners

Super-fast LLM API providers: speed becomes a product feature

What happened
A KDnuggets roundup lists “Top 5 super fast LLM API providers” focused on low-latency and high-throughput inference for open-source models. The providers named are Cerebras, Groq, SambaNova, Fireworks, and Baseten, with comparisons framed around speed-oriented metrics.

Why it matters
Latency is now a user experience constraint, not just an infrastructure detail—especially for voice, copilots, and interactive agents. Speed also changes architecture decisions: some teams may accept higher cost for better responsiveness or higher QPS headroom.

Key details

• The list names Cerebras, Groq, SambaNova, Fireworks, and Baseten.
• The focus is on low latency and high throughput for serving LLMs.
• The framing emphasizes speed metrics (e.g., time-to-first-token and tokens/sec) as decision inputs.

Source links
https://www.kdnuggets.com/top-5-super-fast-llm-api-providers?utm_source=openai

Feature stores: why they’re back (and how “agentic AI” changes the story)

What happened
A KDnuggets explainer revisits feature stores—tracking origins from Uber’s 2017 coining of the term and the subsequent growth of vendor and cloud-native offerings. It argues feature stores are resurging alongside operational AI needs and real-time systems, and cites tools including Feast, Tecton, Vertex AI Feature Store, and SageMaker Feature Store.

Why it matters
As systems shift from one-off model deployments to always-on decisioning, feature quality becomes a reliability layer: freshness, point-in-time correctness, and consistency between training and serving. For agentic systems, the “inputs” are often the hidden failure mode—feature stores aim to make those inputs repeatable and governed.

Key details

• The article notes Uber coined “feature store” in 2017.
• It references vendorization and mentions Tecton (founded 2019).
• Tool examples include Feast, Tecton, Vertex AI Feature Store, and SageMaker Feature Store.

Source links
https://www.kdnuggets.com/all-about-feature-stores?utm_source=openai

Human-in-the-loop (HITL) plan-and-execute agents (LangGraph / LangChain pattern)

What happened
LangChain’s documentation describes human-in-the-loop patterns where an agent can pause (“interrupt”) before executing a tool call, preserve state, and resume after approval or edits. This is presented as a practical way to add checkpoints when agents interact with external systems.

Why it matters
The fastest way to make agents usable in real workflows is controlled autonomy: keep the planning benefits, but insert explicit gates before irreversible actions. HITL also supports auditability—operators can see what the agent intended to do and what was approved.

Key details

• The pattern includes interrupting execution before tool calls.
• State can be saved and the workflow resumed after human input.
• The documented approach is framed for safer operation when tools can affect real systems.

Source links
https://docs.langchain.com/oss/python/langchain/human-in-the-loop?utm_source=openai

JointFM (DataRobot): “digital quant” + zero-shot joint distribution forecasting

What happened
DataRobot published a post describing JointFM, a foundation model for zero-shot joint distributional forecasting in multivariate time series. It claims the model can generate coherent multivariate future scenarios in milliseconds and frames this as enabling near-real-time portfolio optimization workflows.

Why it matters
“Foundation model” ideas are expanding into structured forecasting and scenario generation, where dependencies between variables (correlations and tail behavior) are often the real risk surface. If scenario generation is fast enough, it becomes plausible to integrate into interactive decision loops instead of batch-only analysis.

Key details

• Described as zero-shot joint distributional forecasting for multivariate time series.
• Claims scenario generation in milliseconds.
• Claims it can generate “thousands of coherent future scenarios” quickly.
• States training used an “infinite stream” of synthetic dynamics from stochastic differential equations (SDEs).
• Reports an evaluation summary across 200 controlled synthetic trials matching a classical benchmark on risk-adjusted performance.

Source links
https://www.datarobot.com/blog/instant-portfolio-optimization-with-jointfm/

Alibaba Qwen3.5-397B-A17B: MoE, long context, and agent-readiness

What happened
MarkTechPost reported that Alibaba’s Qwen team released Qwen3.5 and highlighted Qwen3.5-397B-A17B, described as a Mixture-of-Experts (MoE) model with 397B total parameters and ~17B active parameters. Coverage also emphasized long-context capability (262k native context, with references to a 1M-token hosted variant) and agent-oriented positioning.

Why it matters
MoE is a deployment story as much as a modeling story: the goal is large capacity without paying full compute for every token. Long context can reduce retrieval complexity for some workflows (e.g., large docs or logs), while also raising new questions about prompt cost, latency, and evaluation discipline.

Key details

• Model highlighted: Qwen3.5-397B-A17B (397B total parameters, ~17B active via MoE).
• Long-context capability cited as 262k native context; coverage references a 1M-token hosted variant.
• One published configuration lists 60 layers, hidden size 4096, and 512 experts (10 routed + 1 shared active per token).

Source links
https://www.marktechpost.com/2026/02/16/alibaba-qwen-team-releases-qwen3-5-397b-moe-model-with-17b-active-parameters-and-1m-token-context-for-ai-agents/?utm_source=openai
https://huggingface.co/mlx-community/Qwen3.5-397B-A17B-4bit?utm_source=openai

Agoda open-sources “API Agent”: turn REST/GraphQL APIs into an MCP server

What happened
InfoQ reported that Agoda engineers open-sourced “API Agent,” describing it as a universal approach to exposing REST or GraphQL APIs through an MCP server configuration rather than building bespoke MCP servers per API. The report highlights schema introspection, a DuckDB post-processing step to reduce large responses, and security defaults such as read-only mode unless mutations are explicitly enabled.

Why it matters
As MCP-style tooling becomes a common interface, integration work becomes the bottleneck—teams want a safe, repeatable way to connect agents to internal systems. The DuckDB layer is a pragmatic pattern: reduce and reshape data before sending it into model context, improving both cost and reliability.

Key details

• Targets REST and GraphQL APIs, using schema introspection (GraphQL types/fields; REST via OpenAPI specs or examples).
• Uses DuckDB SQL post-processing to manage context size.
• Security defaults include read-only mode; mutations blocked unless explicitly enabled/whitelisted.
• Mentions observability integrations including OpenTelemetry and tracing.

Source links
https://www.infoq.com/news/2026/02/agoda-api-agent/?utm_source=openai

Takeaway: The most durable progress right now isn’t a single model launch—it’s the steady tightening of the whole loop: faster inference, safer agent execution, cleaner integrations, and data infrastructure that keeps real-time systems consistent under pressure.