AI Agents Hit Enterprise Scale — and Expose Infrastructure

Autonomous AI agents moved from pilot curiosity to operational infrastructure this week, as Anthropic reversed a month-old policy on third-party agent access, Intercom rebranded entirely around its AI agent product, and enterprise security researchers warned that identity governance frameworks are failing to keep pace with non-human workloads running in production.

Anthropic Reinstates Third-Party Agent Access — With Metered Credits

Anthropic reversed course on its April 2026 ban on third-party agent usage, announcing via its @ClaudeDevs account that all paid subscribers will now receive a dedicated pool of “Agent SDK” credits usable for programmatic and external agent workflows — including the open-source OpenClaw harness that triggered the original crackdown.

The policy shift comes roughly a month after Anthropic prohibited Claude Pro and Max subscribers from routing their subscriptions through autonomous agents. The core problem, as VentureBeat reported, was financial: subscribers paying $20 to $200 per month were consuming hundreds or thousands of dollars in token volume through agents, creating an unsustainable gap between subscription revenue and actual compute costs.

The new Agent SDK credit system separates interactive usage — Claude.ai chat, Claude Code terminal sessions — from programmatic API-style calls made by external agents. Interactive usage continues drawing from standard subscription limits. Agent SDK credits form a distinct, metered bucket.

Developer reaction was skeptical. The backlash centers on the gap between what users previously consumed effectively for free under their subscriptions and what the new metered credits actually provide. Anthropic has not published the conversion rate between Agent SDK credits and token volume, leaving builders uncertain about real-world cost implications before committing workflows to the new system.

Enterprise IAM Was Not Built for Non-Human Agents

While Anthropic works out subscription economics, enterprises deploying agents at scale face a more structural problem: identity and access management systems designed for human employees are breaking down under the weight of non-human agent identities.

Cisco President Jeetu Patel told VentureBeat at RSAC 2026 that 85% of enterprises are running agent pilots, but only 5% have reached production — an 80-point gap he attributed directly to trust and governance deficits, not model capability or compute availability.

The stakes are concrete. A medical transcription agent updating electronic health records in real time and a computer vision system running quality control on a manufacturing line both generate non-human identities that most enterprises cannot inventory, scope, or revoke at machine speed. IANS Research found that most businesses lack role-based access control mature enough for their existing human identities — agents compound that deficit significantly.

The threat surface is already widening. The 2026 IBM X-Force Threat Intelligence Index reported a 44% increase in attacks exploiting public-facing applications, driven by missing authentication controls and AI-enabled vulnerability discovery.

Michael Dickman, SVP and GM of Cisco’s Campus Networking business, described the problem to VentureBeat as architectural rather than a tooling gap: enterprises need a trust framework that assigns scope, accountability, and revocation rights to agents before they touch sensitive systems — not after an incident forces the issue.

Intercom Renames Itself Fin, Launches Agent-for-Agents

The company formerly known as Intercom formalized a bet that AI agents are not a product feature but the entire business. CEO Eoghan McCabe renamed the 15-year-old company to Fin two days before the company launched Fin Operator at a live event in San Francisco Thursday.

Fin Operator is an AI agent whose sole job is managing Fin, the company’s customer-facing AI agent. Rather than automating customer interactions — which Fin handles on the front end — Operator targets support operations teams: the people updating knowledge bases, debugging conversation failures, and monitoring performance dashboards. Brian Donohue, VP of Product, described the distinction plainly to VentureBeat: “Fin is an agent for your customers. Operator is an agent for your support ops team.”

The financial context makes the rebrand legible. Fin has crossed $100 million in annual recurring revenue and is growing at 3.5x. The broader company generates $400 million in ARR, meaning the AI agent product now accounts for roughly a quarter of total revenue and nearly all of its growth trajectory. Fin Operator enters early access for Pro-tier users immediately, with general availability planned for summer 2026.

The product architecture — one agent supervising another — reflects a pattern emerging across enterprise AI deployments: as front-line agents multiply, the operational burden of managing them creates demand for a second layer of automation.

Hardware Demand Scales With Agentic Workloads

The infrastructure economics behind agent deployment surfaced at Dell Technologies World, where NVIDIA CEO Jensen Huang joined Dell Chairman and CEO Michael Dell on stage to frame the compute requirements of agentic AI.

Dell sized the market: worldwide AI infrastructure spending could reach $3–4 trillion by 2030, with token consumption projected to grow 3,400% in the same window. “The rate of change has gone parabolic, and it’s not slowing down,” Dell said, according to the NVIDIA AI Blog.

Huang’s pitch centered on cost reduction at scale. The NVIDIA Vera Rubin NVL72 delivers agentic AI inference at one-tenth the cost per token compared to prior generations. Agent sandboxes run 50% faster on the NVIDIA Vera CPU than on traditional CPUs, and enterprise data queries run up to 3x faster. NVIDIA cited 5,000 enterprises — including Lilly, Samsung, and Honeywell — running AI workloads on Dell AI Factories with NVIDIA hardware.

The cost-per-token figure is directly relevant to the subscription economics Anthropic is navigating. As inference costs fall, the gap between flat-rate subscription pricing and actual token consumption becomes more manageable — but only if hardware efficiency gains translate into provider pricing before subscriber expectations calcify around the old model.

Empromptu Targets Training Data Waste in Production Workflows

San Francisco-based Empromptu AI launched Alchemy Models Thursday with a premise aimed at a different layer of the agent stack: the training signal that production workflows generate and most enterprises discard.

Every query an enterprise AI application processes, every correction a subject matter expert makes to its output, is potential fine-tuning data. According to VentureBeat, Alchemy captures that signal automatically, routing validated expert outputs into a continuous fine-tuning pipeline. Enterprises own the resulting model weights outright.

The positioning sits between retrieval-augmented generation and traditional fine-tuning. RAG retrieves external context at inference without modifying weights. Traditional fine-tuning modifies weights but requires separately assembled labeled datasets and a dedicated ML pipeline. Alchemy runs the latter continuously, using the live enterprise application as its data source — no ML team required to initiate or maintain the loop.

Empromptu CEO Shanea Leven told VentureBeat that the constraints driving interest are consistent across customers: inference costs that scale with usage, no ownership of the models their data is effectively training, and limited ability to customize behavior for domain-specific tasks.

What This Means

This week’s developments collectively describe an agent ecosystem that has outrun the infrastructure built to support it — on three distinct fronts.

On the pricing side, Anthropic’s reversal on OpenClaw access reveals that flat-rate subscriptions and autonomous agents are fundamentally incompatible at scale. The Agent SDK credit system is a reasonable architectural fix, but the metered reality it introduces will reshape which developer workflows remain economically viable on subscription tiers versus direct API access. Builders who built cost assumptions around subscription pricing will need to recalculate.

On the security side, the 80-point gap between pilot and production identified by Cisco is not a temporary lag — it reflects a genuine architectural deficit in enterprise IAM. Non-human agent identities require scoping, auditing, and revocation capabilities that most access control systems were not designed to provide. Until that gap closes, high-stakes deployments in healthcare and manufacturing will remain bottlenecked by governance concerns rather than model performance.

On the economics side, NVIDIA’s cost-per-token improvements and Empromptu’s continuous fine-tuning approach both point toward the same structural shift: the value in enterprise AI is migrating from model access toward ownership of customized, domain-specific weights trained on proprietary workflows. Organizations that treat their production interactions as disposable are effectively subsidizing foundation model providers’ training pipelines without capturing the resulting value themselves.

The agent-managing-agent pattern Fin Operator represents may be the clearest signal of where the market is heading: as front-line agents scale, the operational complexity of managing them becomes a product category in its own right.

FAQ

What are Anthropic’s new Agent SDK credits?

Anthropic introduced Agent SDK credits as a separate usage bucket within paid Claude subscriptions (Pro and Max tiers), specifically for programmatic and third-party agent workflows like OpenClaw. They are distinct from standard interactive usage limits that apply to Claude.ai chat and Claude Code sessions.

Why are enterprises struggling to move AI agents from pilot to production?

According to Cisco President Jeetu Patel, 85% of enterprises are running agent pilots but only 5% have reached production — a gap driven primarily by identity governance deficits. Most enterprise IAM systems cannot inventory, scope, or revoke non-human agent identities at the speed and scale that production deployments require.

What is Fin Operator and how does it differ from Fin?

Fin is the customer-facing AI agent built by the company formerly known as Intercom, now rebranded as Fin. Fin Operator is a separate AI agent designed to manage Fin itself — handling tasks like knowledge base updates, conversation debugging, and performance monitoring for the support operations teams that run Fin day-to-day.