The Stack Observer
Agentic AI / AI

The Week Agentic AI Became the Default

•Stackxx•Agentic AI, AI

Agentic AI is no longer a research curiosity. Over the past two weeks, every major player in the AI ecosystem has made a move that shifts the conversation from what models can do to what agents can accomplish. Google announced an operating system for autonomous intelligence. DeepSeek shipped a model purpose-built for long-running tool workflows. IBM launched the first open benchmark that evaluates complete agent systems, not just the models inside them. And NVIDIA, LangChain, and Ollama all released infrastructure that makes building and deploying agents measurably easier.

Here is what changed, and what it means for anyone building with AI.

Google I/O 2026: An Operating System for Agents

Google I/O 2026 was not about a single model. It was about turning the entire Google stack into an agent platform. Sundar Pichai and the Google AI team announced Gemini 3.5 Flash, Gemini Omni, Google Antigravity, and — most significantly — information agents embedded directly into Search.

Gemini 3.5 Flash is Google’s first explicitly agent-first model. It scores 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, and 83.6% on MCP Atlas — benchmarks designed for agentic reasoning, not just chat. Google positioned it as the default model for Search’s AI Mode, which now serves over one billion users monthly. The implication is clear: Google is betting that the next interface for Search is not a query box, but an agent that operates continuously on your behalf.

The information agents are the most radical part. These are persistent background agents that monitor the web — blogs, news, social posts, financial data, shopping signals — and synthesize updates for specific user-defined tasks. You can run multiple agents simultaneously. They send intelligent, synthesized updates and can take action. Google is rolling this out first to Pro and Ultra subscribers this summer.

Google Antigravity is the agent-first development platform behind it all. It powers the generative UI in Search — custom layouts, tables, graphs, and even full mini-apps generated on the fly for a specific query. Combined with Universal Cart (an intelligent shopping agent that tracks prices, compatibility, and deals across the web), Google is building an end-to-end agentic layer on top of its existing services.

This is not just product expansion. It is a re-architecture of how users interact with information.

DeepSeek V4: A Model Designed for Agents, Not Chat

While Google is building the platform, DeepSeek is building the engine. DeepSeek-V4 launched with two variants: V4-Pro at 1.6T parameters (49B active) and V4-Flash at 284B parameters (13B active). Both ship with a one-million-token context window.

The headline is not the parameter count. It is the architecture. V4 introduces Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), which reduce per-token inference FLOPs by up to 90% and KV cache memory to roughly 2% of what standard grouped-query attention requires. At one million tokens, V4-Pro uses 27% of the FLOPs that DeepSeek-V3.2 would need. This is not incremental improvement. It is the difference between a model that can run long contexts and one that should.

DeepSeek built this for agentic workloads. Long-running tool-use trajectories — multi-step browsing sessions, terminal command chains, SWE-bench tasks — blow past context budgets and fill GPU memory. V4 fixes both problems at the architecture level.

The post-training decisions are equally deliberate. V4 preserves reasoning content across user message boundaries when tool calls are present, so an agent retains its accumulated state across multi-turn workflows. It introduces a |DSML| special token and an XML-based tool-call format to reduce the escaping failures common in JSON-in-string tool calls. And it was trained in DSec, a Rust sandbox infrastructure capable of running hundreds of thousands of concurrent RL rollouts across function calls, containers, microVMs, and full VMs.

The benchmark numbers support the design. V4-Pro-Max scores 80.6 on SWE Verified (tied with Gemini 3.1 Pro), 73.6 on MCPAtlas Public (second only to Claude Opus 4.6 Max), and 51.8 on Toolathlon. In DeepSeek’s internal R&D coding benchmark, it hits 67% pass rate versus 47% for Claude Sonnet 4.5. In a survey of 85 DeepSeek developers using it as their daily driver, 52% said it was ready to replace their current primary coding model.

DeepSeek V4 is not trying to be the best chat model. It is trying to be the best agent model. That distinction matters.

The Open Agent Leaderboard: Measuring Systems, Not Models

IBM Research and Hugging Face launched the Open Agent Leaderboard, the first open benchmark that evaluates complete agent systems rather than isolated models. It reports both quality and cost across six diverse benchmarks: SWE-Bench Verified, BrowseComp+, AppWorld, tau2-Bench Airline & Retail, tau2-Bench Telecom, and a unified research task.

The leaderboard is paired with Exgentic, an open evaluation framework that standardizes task definitions, contexts, and action protocols across benchmarks. This matters because changing the tools, planning strategy, memory system, or error recovery mechanism around the same model can produce radically different results. The leaderboard makes that visible.

One early finding is already reshaping how teams think about agent procurement: general-purpose agents are competitive with specialized ones. In several cases, agents with no benchmark-specific tuning matched systems built directly for those tasks. Tool shortlisting — helping the agent focus on relevant tools instead of searching through everything — improved performance across every model tested and turned failing configurations into viable ones.

Another finding is about failure economics. Failed agent runs cost 20–54% more than successful ones. For production deployments, how an agent fails is as important as how often it succeeds. The leaderboard captures both dimensions.

The leaderboard currently includes five models across five agents. Open-weight models (DeepSeek V3.2 and Kimi K2.5) are competitive on specific combinations but trail frontier closed-source models by 18–29 percentage points on average. The gap is real, but it is also closing.

The Agent Customization Stack

NVIDIA published a comprehensive guide to agent customization, outlining nine techniques from prompt engineering to reinforcement learning. The techniques range in complexity: system prompts and RAG for quick wins; tool injection and supervised fine-tuning for domain specialization; and RL and distillation for fundamental behavior change.

NVIDIA is also building infrastructure. NVIDIA-Verified Agent Skills provide capability governance — a standardized way to declare what an agent can and cannot do, with cryptographic verification. This is aimed at enterprises that need to deploy agents with auditable constraints. The NVIDIA Dynamo inference stack now includes optimizations specifically for agentic workloads, where non-deterministic trajectories (actions, observations, tool calls) break the assumptions of traditional batch inference.

On the open-source side, Ollama is in pre-release for v0.30.0, which re-architects the runtime to directly support llama.cpp instead of building on GGML, with MLX acceleration on Apple Silicon. LangChain shipped langchain-tests 1.1.9 with improved streaming assertions and audio chat integration tests. OpenClaw released gateway performance improvements (plugin metadata caching, lazy-load handlers, process-stable channel catalogs) and added real-time voice consult steering — the ability to ask for agent run status, cancel, or queue follow-up work during an active consult.

These are not isolated updates. They are converging on the same problem: making agents fast, observable, and safe enough to run in production.

What This Means for Builders

Three shifts are now in motion simultaneously.

First, the interface is becoming the agent. Google’s information agents in Search, OpenClaw’s real-time consult steering, and Anthropic’s Claude Design all point to the same trend: users will interact with AI through persistent, stateful agents rather than one-off chat sessions. The browser tab, the terminal, and the search box are being replaced by the agent runtime.

Second, models are being optimized for agentic workloads, not chat. DeepSeek V4’s architecture, Google’s Gemini 3.5 Flash benchmarks, and the Open Agent Leaderboard’s emphasis on full-system evaluation all reflect a move away from measuring perplexity and toward measuring task completion over long horizons. The community is beginning to treat context efficiency, tool reliability, and failure cost as first-class metrics.

Third, evaluation is catching up to capability. The Open Agent Leaderboard, NVIDIA-Verified Agent Skills, and Braintrust’s emphasis on evals as the new PRD all signal that the industry is maturing. Building an agent is no longer enough. You need to prove it works, prove it is safe, and prove it is cost-effective.

The agentic era is not coming. It is here. The only question is whether your stack is ready for it.

Sources

Next signal