AI Agents Hit Enterprise Scale—and Five New Fault Lines

Autonomous AI agents moved from lab curiosity to operational infrastructure this week, with five concurrent developments exposing where the technology is maturing and where it is breaking down: Anthropic reversed a month-old ban on third-party agent access, Exaforce closed a $125 million Series B for agentic security operations, Intercom rebranded entirely around its AI agent and launched a second agent to manage the first, Empromptu AI released a platform that trains custom models from live workflow data, and a VentureBeat analysis found that 85% of enterprises are running agent pilots while only 5% have reached production.

Anthropic Reverses OpenClaw Ban with New “Agent SDK” Credits

Anthropie reversed course on a policy it introduced in early April 2026 that blocked Claude paid subscribers from using their subscriptions to power third-party agentic tools like OpenClaw, the popular open-source autonomous agent harness. According to an announcement from Anthropic’s official developer account @ClaudeDevs, all paid subscribers now receive a new subcategory of “Agent SDK” credits that can be allocated specifically for programmatic and third-party agent use.

The original ban stemmed from a financial mismatch: subscribers paying $20 to $200 per month under Claude Pro and Max plans were consuming hundreds or thousands of dollars in tokens through autonomous agents, an unsustainable gap for Anthropic’s compute infrastructure. Even during the prohibition, Anthropic never fully disabled the technical capability — it redirected users toward its paid API instead.

The new credit system attempts to meter that usage rather than block it outright. Standard interactive use — chatting in the browser or running Claude Code in a terminal — continues drawing from the main subscription pool. Programmatic and agent-driven calls now draw from the separate Agent SDK allocation, according to Anthropic technical staffer Lydia Hallie in a post clarifying the policy.

The developer community’s reaction was mixed. The reversal was welcomed in principle, but the metered reality drew criticism from builders who had grown accustomed to effectively unlimited agentic throughput at flat subscription prices. The credits model represents a structural shift: agent usage is no longer bundled into a flat subscription but rationed against a separate, finite pool.

Identity Governance Is the Real Bottleneck for Enterprise Agents

The gap between pilot and production for enterprise AI agents is not a model capability problem. It is an identity problem.

Cisco President Jeetu Patel told VentureBeat at RSAC 2026 that 85% of enterprises are running agent pilots while only 5% have reached production — an 80-point gap he attributed directly to trust and governance deficits. The core issue: agents generate non-human identities that most enterprise identity and access management (IAM) systems cannot inventory, scope, or revoke at machine speed.

IANS Research found that most businesses still lack role-based access control mature enough for their existing human identities — agents compound that problem significantly. 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.

The practical stakes are concrete. A medical transcription agent updating electronic health records in real time, or a computer vision agent running quality control on a manufacturing line, both generate non-human identities with privileged access to sensitive systems. When something goes wrong, most enterprises currently have no clear answer to two questions every CISO asks first: which agents have production access, and who is accountable?

Michael Dickman, SVP and GM of Cisco’s Campus Networking business, described the problem as architectural rather than a tooling gap — existing IAM frameworks were designed for human users authenticating at human speeds, not for agents that may spin up, act, and terminate within seconds.

Exaforce Raises $125M to Put Agents Inside Security Operations

Exaforce closed a $125 million Series B on Tuesday, bringing its total funding to $200 million, to expand its agentic security operations center (SOC) platform. The round was led by HarbourVest, with participation from Peak XV, Mayfield, Khosla Ventures, Seligman Ventures, and AICONIC, according to SecurityWeek.

The platform centers on autonomous agents called Exabots that handle the full security operations lifecycle:

• Detection and triage of incoming alerts
• Investigation using natural language search over live security context
• Response actions including user verification and access revocation
• Correlation of high-volume telemetry without reliance on traditional SIEM rules or manual query languages

A real-time knowledge graph connects events, identities, permissions, configurations, and activity at ingest time. The system combines data semantics, machine learning, behavioral baselining, and large language models into a multi-model AI engine.

“We built Exaforce to be the platform defenders actually work in, not just an AI layer on top of existing tools,” Ankur Singla, CEO of Exaforce, said in the announcement. Exaforce will use the funding to expand into Japan and Europe.

The Exaforce raise is notable precisely because it sits at the intersection of the identity governance problem described above: an agentic platform operating inside security infrastructure, where the consequences of misconfigured agent permissions are severe.

Intercom Becomes Fin, Then Builds an Agent to Manage Its Agent

The company formerly known as Intercom took two significant steps this week. CEO Eoghan McCabe formally renamed the 15-year-old company to Fin — a signal that the AI agent product is now the core business rather than a feature of it. Two days later, at a live event in San Francisco, the company launched Fin Operator, an AI-powered system designed to manage Fin itself.

Fin, the customer-facing agent, recently crossed $100 million in annual recurring revenue and is growing at 3.5x. The broader company generates $400 million in ARR, meaning the agent accounts for roughly a quarter of total revenue and virtually all growth momentum.

Fin Operator targets a different user: the support operations teams who configure, monitor, and improve Fin’s behavior — updating knowledge bases, debugging conversation failures, and reviewing performance dashboards. Brian Donohue, VP of Product, described the distinction to VentureBeat: “Fin is an agent for your customers. Operator is an agent for your support ops team.”

This is a meaningful architectural choice. Rather than expecting human operators to manage agent behavior through dashboards and manual configuration, Fin is deploying a second agent to handle that management layer. Fin Operator enters early access for Pro-tier users immediately, with general availability planned for summer 2026.

The agent-manages-agent pattern is likely to recur across enterprise software categories as the complexity of maintaining production AI agents outpaces what human operators can handle manually.

Empromptu AI Launches Continuous Fine-Tuning from Live Workflows

San Francisco-based Empromptu AI launched Alchemy Models on Thursday with a specific claim: enterprises are generating continuous training signal from their production AI applications and discarding it.

Alchemy captures validated outputs from subject matter expert corrections within running applications and routes them back into a fine-tuning pipeline, updating model weights continuously. Enterprises own the resulting model weights outright — a meaningful distinction from foundation model API arrangements where usage data trains models the customer does not own.

The platform occupies different territory from both retrieval-augmented generation (RAG) and traditional fine-tuning:

• RAG retrieves external context at inference time without modifying weights
• Traditional fine-tuning modifies weights but requires separately assembled labeled datasets and a dedicated ML pipeline
• Alchemy modifies weights continuously, using the enterprise application itself as the data source

Empromptu CEO Shanea Leven described the underlying concern driving adoption: enterprises face compounding constraints from inference costs that scale with usage, no ownership of models their data is effectively training, and limited ability to customize behavior for domain-specific tasks. Alchemy targets all three simultaneously by converting workflow interactions into proprietary model improvements.

What This Means

Taken together, this week’s developments sketch a clear picture of where agentic AI stands in mid-2026: the technology is capable enough to run hospital records systems and factory inspections, but the infrastructure around it — identity governance, access control, cost metering, and agent management — has not kept pace.

Anthropice’s Agent SDK credit reversal illustrates the economic tension at the subscription layer: flat-rate pricing cannot survive unlimited agentic throughput, and the industry is converging on metered models that more closely mirror API pricing. That shift will reshape the economics of developer-facing AI products.

The 80-point gap between enterprise pilot and production — 85% piloting, 5% in production per Cisco’s Jeetu Patel — is the most consequential number in this set of stories. It suggests that model capability is no longer the primary constraint on enterprise AI adoption. Governance, accountability, and identity management are. Exaforce’s $125 million raise and the VentureBeat IAM analysis both point toward a significant infrastructure buildout in agentic security and identity over the next 12-18 months.

Fin’s decision to build an agent to manage its agent is an early signal of a structural pattern: as agent deployments scale, the operational overhead of managing them manually becomes a bottleneck, and the solution is more agents. That recursion will raise its own governance questions.

Empromptu’s Alchemy sits at the intersection of two pressures: enterprises want to reduce inference costs and own their model improvements, and the continuous fine-tuning approach addresses both without requiring a dedicated ML team. If the approach scales, it could shift the fine-tuning market away from one-time, dataset-driven projects toward continuous, workflow-integrated model improvement.

FAQ

What are Agent SDK credits on Claude subscriptions?

Agent SDK credits are a new subcategory of usage allocation that Anthropic introduced for all paid Claude subscribers in May 2026. They are separate from standard subscription limits and can be used specifically for programmatic and third-party agent access, including tools like OpenClaw, rather than drawing from the same pool as interactive chat usage.

Why are only 5% of enterprises running AI agents in production?

According to Cisco President Jeetu Patel, speaking at RSAC 2026, the primary barrier is identity governance — most enterprise IAM systems were built for human users and cannot inventory, scope, or revoke the non-human identities that AI agents generate at machine speed. IANS Research and the 2026 IBM X-Force report both identify missing authentication controls as a compounding factor.

What does Fin Operator do differently from Fin itself?

Fin is a customer-facing AI agent that handles support interactions. Fin Operator is a separate AI system that manages Fin — handling tasks like updating knowledge bases, diagnosing conversation failures, and monitoring performance dashboards. It targets support operations teams rather than end customers, and enters early access for Pro-tier users in summer 2026.