The Agentic AI Moment Has Arrived

For years, the AI conversation has been about models, benchmarks, and token costs. That era is ending. In its place, a more practical question is taking hold: can an AI agent actually do useful work at scale, and can we measure whether it does?

Over the past two weeks, a wave of announcements from OpenAI, Google, Anthropic, and enterprise adopters like Cars24 has made the answer increasingly clear. The agentic era is not a future possibility. It is a present reality, and the companies figuring out how to govern, measure, and scale it are pulling ahead of those still treating AI as a chatbot upgrade.

OpenAI Pushes Agentic Economics into the Boardroom

OpenAI has spent the past week making a deliberate argument to CFOs and enterprise leaders: stop measuring AI by adoption and start measuring it by work accomplished. In a post authored by CFO Sarah Friar, the company introduced a practical scorecard it calls “Useful Intelligence per Dollar.” The idea is simple but underutilized. Rather than obsessing over cost per token, enterprises should track four questions: Is AI completing work that matters? What does each successful task cost? Can people depend on the result? And does each AI dollar produce more value as usage grows?

The timing is not accidental. GPT-5.6, OpenAI’s latest flagship model series, launched with a three-tier architecture: Sol for frontier reasoning, Terra for balanced performance and cost, and Luna for fast, high-volume tasks. On the Artificial Analysis Coding Agent Index, GPT-5.6 Sol reached a new state of the art while using 54% fewer output tokens and 57% less time per task than competing models. On DeepSWE v1.1, which evaluates long-horizon engineering tasks, GPT-5.6 Sol scored 72.7%, outperforming Claude Fable 5 at 69.9%.

OpenAI is also making GPT-5.6 the preferred model across Microsoft 365 Copilot, meaning the same improvements in reasoning and token efficiency will now flow into Word, Excel, PowerPoint, and Copilot Chat for millions of enterprise users. That partnership shift signals a broader trend: the model war is becoming an integration war. Winning model performance is only part of the equation. The rest is getting that performance into workflows people already use.

From Chat to Work: How Cars24 Built an AI-First Operating Model

While vendors race on benchmarks, real-world adopters are writing the playbook for how agentic AI actually runs inside a business. Cars24, one of India’s largest used-car marketplaces, offers perhaps the clearest example of what happens when agents stop being demos and start becoming infrastructure.

The company deployed OpenAI-powered voice and chat agents across the entire customer lifecycle: discovery, financing, test drives, document collection, follow-ups, post-purchase service, and seller re-engagement. The agents do not just answer questions. They recommend vehicles, book appointments, reschedule missed inspections, collect financing details, and recover leads that used to drop out after ten days.

The numbers are instructive. Cars24 now handles over one million conversation minutes per month through AI agents. Critically, the company also deployed Codex and ChatGPT Enterprise to roughly 600 employees, with 85% to 90% daily active usage. Product managers use Codex to create and refine Linear tickets. Engineering teams tag Codex into bug reports so it can pick up defined tasks and summarize work across GitHub. Finance teams use it to pull numbers from systems of record, run analysis, and prepare investor reports. Purchase order reviews above a defined threshold are now partially automated, with Codex flagging anomalies and auto-approving clean requests.

As Jayesh Gupta, the company’s head of AI and Innovation, noted, the shift went beyond engineering. “I thought Codex would make our engineers faster. What surprised me was how quickly it spread beyond engineering. Product managers, finance teams, and even day-to-day workflows started changing.”

This is the pattern that matters. Agentic AI succeeds not when a single team pilots it, but when it becomes an operating layer that multiple functions build on top of.

Google Expands Managed Agents with Background Execution and Remote MCP

On the infrastructure side, Google is expanding what its Gemini API can do for developers building production agents. In early July, the company announced new capabilities for Managed Agents in the Gemini Interactions API, including background execution, remote MCP server integration, custom function calling, and credential refresh.

Background execution is particularly important. Instead of holding an HTTP connection open for long-running tasks, developers can pass a background flag and receive an interaction ID. The agent runs asynchronously in Google’s isolated cloud sandbox while the client polls or reconnects later. This addresses a real fragility in current agent architectures: many agent tasks require minutes or hours of reasoning, code execution, and tool use, and keeping a live connection open for that duration is brittle at best.

Remote MCP server integration is another meaningful addition. The Model Context Protocol, an open standard for connecting AI agents to external tools and data, has gained significant traction this year. Google’s new support allows managed agents to connect directly to remote MCP servers alongside built-in sandbox tools like Google Search and code execution. This means a developer can give an agent access to internal databases, observability systems, or proprietary APIs without writing custom proxy middleware.

Google’s direction here is clear: make the Gemini API a reliable substrate for agents that need to operate across enterprise systems, not just answer questions from public knowledge.

Together, background execution and remote MCP support address two of the most common reasons agent projects stall in production: timeouts on long tasks and the complexity of bridging internal systems. For developers, this lowers the barrier to building agents that span multiple services and run for extended periods, which is exactly what enterprise workflows typically require.

Anthropic Returns Fable 5 and Ships Claude Science

Anthropic’s announcements over the past month have been more focused but no less significant. Fable 5, the company’s coding-focused model, redeployed globally on July 1 after a period of restricted availability. Anthropic also joined Amazon, Microsoft, Google, and other Glasswing partners in proposing an industry-wide framework for scoring jailbreak severity, an acknowledgment that as agents gain more capability, the security community needs common vocabulary and measurement.

More distinctively, Anthropic launched Claude Science, an AI workbench designed for researchers. It integrates the tools and packages scientists most often use, produces auditable artifacts, and provides flexible access to computing resources. Unlike general-purpose coding agents, Claude Science is built around reproducibility and attribution: the kind of trustworthiness that matters when AI is generating analysis that could inform published research or policy decisions.

The launch of Claude Science also reflects a growing divergence in the agent market. One path optimizes for speed and breadth, exemplified by Codex and general coding agents that aim to accelerate every software task. The other path, which Anthropic is pursuing, optimizes for trust and auditability, targeting domains where a wrong answer is not just inconvenient but consequential. Both paths are valid, and both are growing. The difference is which tradeoffs each team is willing to make.

What Building Production Agents Actually Takes

For all the vendor announcements, the most sobering perspective on agentic AI came from AI2’s Skylight team, which built Shippy, a maritime domain awareness agent used by analysts to track fishing vessels and protected areas. In a detailed technical writeup, the team described what separates a prototype from something you can trust in a high-stakes environment.

Shippy’s architecture breaks an agent into three parts: a soul (the system prompt that sets behavioral boundaries), skills (versioned markdown files defining how to handle specific tasks), and config (runtime settings like model choice and agent harness). The critical insight is that while the model itself is nondeterministic, the tools it reaches for should not be. Instead of letting the agent construct raw API calls, the Skylight team built a purpose-built CLI that collapses API complexity into predictable commands. This eliminates a steady stream of subtle bugs: malformed pagination, geometry encoding errors, and filter type mismatches that produced correct-looking but wrong data.

The lesson generalizes. As teams move from chat to agents that take action across systems, the failure mode shifts from “the model gave a wrong answer” to “the model took a wrong action on a real system.” Trustworthy agents require deterministic tool boundaries, explicit governance over what data and systems the agent can touch, and clear escalation paths when the agent exceeds its competence.

The New Benchmark: Useful Work, Not Just Smarter Models

Agentic AI in 2026 is entering a new phase. The frontier models are now close enough in capability that raw benchmark leadership matters less than integration, governance, and ROI measurement. As OpenAI’s scorecard framework suggests, the winners will be the organizations that can answer four questions honestly: How much useful work gets done? What does a successful task actually cost? How often does AI get the work right? And does each AI dollar buy more work as usage grows?

The companies that have moved fastest are not necessarily the ones with the biggest models. They are the ones that tied agentic workflows to specific business processes, built deterministic tool boundaries, and treated governance as an operating layer rather than an afterthought. Cars24, Deutsche Telekom, and the early adopters of Google’s Managed Agents are building that muscle now. The rest of the market is about to follow.