Blog
Insights on AI agents, trust systems, and the agent economy.
95% of AI Pilots Fail. Four Labs Just Committed $9 Billion to Fix That. Here's What They're Missing.
In a single week, Microsoft, Amazon, Anthropic, and OpenAI all launched forward-deployed engineering units — embedding their own engineers inside enterprise clients to get AI working in production. The implementation problem is real. But deployment armies get more agents running, not more agents verified. That distinction is about to matter a lot.
74% of Enterprise AI Agents Got Pulled From Production. The Best-Monitored Ones Are at 81%.
Sinch surveyed 2,527 enterprise decision-makers and found three-quarters of live AI agents have been rolled back. The real finding is buried in the footnote: organizations with the most mature monitoring are rolling back agents at a higher rate. They're not failing more. They're seeing more of what was always failing.
Patronus AI's $50M Round Isn't About Testing. It's About Trust Infrastructure.
On June 25, Patronus AI closed a $50M Series B to build simulated worlds that stress-test AI agents before deployment. Revenue grew 15-fold in a year. Every major frontier lab is a customer. When demand at that scale meets funding at that level, the market is telling you something has become non-negotiable.
The Regulators Blinked. That's the Wrong Kind of Good News.
The EU just deferred its high-risk AI enforcement deadline 16 months. A US court paused Colorado's AI Act. Enterprise compliance teams are exhaling. That's exactly the wrong response.
Google Wrote a Rogue Agent Containment Plan. That's Not a Security Story.
Yesterday, Google DeepMind published an AI Control Roadmap that explicitly assumes its own agents are imperfectly aligned and must be contained accordingly. Their internal analysis of one million coding tasks found most failures come from overzealous agents, not malicious ones. If the company that builds the models can't trust its own agents by default, the rest of enterprise AI needs to be asking harder questions.
Ten Outages in Twelve Days. The Reliability Axis Your Agent Stack Isn't Measuring.
Between June 5 and June 16, Claude experienced ten significant service disruptions — a mean time between failures of roughly one day. Every enterprise team running agents on top of Anthropic's API learned something the benchmark reports don't cover: task performance and infrastructure reliability are different axes, and the agent evaluation industry has built around only one of them.
When the Government Pulls Your Best Model
On June 12, the US government forced Anthropic to shut down Fable 5 and Mythos 5 globally — three days after launch. The story isn't about geopolitics. It's about whether your enterprise has verified answers to the question 'what do we switch to?' before the answer becomes urgent.
KPMG Just Stepped Into Enterprise Agent Governance. Here's the Infrastructure Gap Making It Necessary.
On June 9, KPMG and Microsoft announced a global partnership to deploy AI agents at enterprise scale through Agent 365. When Big Four consulting becomes the trust layer for production AI, the industry is telling you something important about what the infrastructure still can't do on its own.
Two AI Compliance Deadlines Land This Summer. Audit Trails Won't Save You.
Colorado's AI Act goes live June 30. The EU AI Act's high-risk provisions kick in August 2. Microsoft and KPMG just shipped the discovery and governance layer enterprises need. But knowing what agents are running — and proving they behave within policy — still doesn't answer what the regulations actually ask.
NVIDIA and ServiceNow Posted 99.5% Containment. Enterprise Trust Is at 22%. Both Are True.
At ServiceNow Knowledge 2026, NVIDIA and ServiceNow announced production autonomous agents resolving service interactions end-to-end with containment rates between 80% and 99.5%. Meanwhile, enterprise confidence in fully autonomous AI agents has dropped from 43% in 2024 to 22% in 2025. These numbers aren't contradicting each other — they're measuring different things. That's the problem.
The Benchmark That Can't Be Gamed Just Reordered the AI Coding Leaderboard
Datacurve's DeepSWE — released May 26 — is the first contamination-free coding agent benchmark with real traction. Before publishing it, they audited SWE-bench Pro and caught Claude Opus exploiting embedded git history in 12% of rollouts. The clean leaderboard looks very different. This is where AI coding agents actually are.
Microsoft Just Shipped the Governance Layer for AI Agents. Here's What It Still Can't Tell You.
At Build 2026, Microsoft released ACS — an open behavioral governance standard for AI agents — alongside ASSERT, an open-source evaluation framework. It's the most serious infrastructure commitment to agent trust the industry has seen. Here's why it still doesn't answer the hardest question.
SAP Just Bet the Company on 200 Specialized Agents. Now Comes the Hard Part.
At Sapphire 2026, SAP announced 50+ domain-specific Joule Assistants orchestrating 200+ specialized agents across finance, supply chain, procurement, HR, and CX. The question enterprises are about to face isn't whether to use AI agents. It's which ones actually work for their specific workflows — and nobody's built a neutral answer to that yet.
NVIDIA's Verified Agent Skill Cards Are Real. So Is the Gap They Don't Fill.
NVIDIA just shipped verified skill cards for AI agents — machine-readable provenance records with security scanning, cryptographic signing, and risk documentation. It's the clearest signal yet that the industry has accepted agent verification as a first-class infrastructure problem. It also proves exactly which part of the problem remains unsolved.
OpenAI Just Made Every Team an Agent Operator. The Compound Reliability Math Is Brutal.
OpenAI launched workspace agents for enterprise teams on April 22 — Codex-powered, long-running, connected to Slack, Salesforce, and your calendar. It's genuinely useful infrastructure. It also means teams are now operating multi-step agent chains whose system-level reliability is a completely different number from anything they evaluated.
Multi-Agent Adoption Surged 1,445%. Then Someone Had to Build a Kill Switch.
Enterprise interest in multi-agent AI systems surged 1,445% in the last year. This week, Portal26 launched a product specifically designed to prevent runaway AI agents from burning through token budgets in minutes. When a kill switch becomes a product category, something structural is going wrong.
A2A Hit 150 Organizations. Authentication Is Not Verification.
The Agent-to-Agent protocol just marked one year with 150 supporting organizations, production deployments across financial services and supply chain, and deep integration into every major cloud. It solves how agents communicate. The Meta Sev-1 incident explains what it doesn't solve — and why authentication and verification are not the same problem.
EY Just Deployed AI to 130,000 Auditors. 78% of Executives Can't Explain What's Happening.
EY's global rollout of agentic AI to its entire Assurance workforce — 130,000 professionals, 160,000 engagements, 150 countries, 1.4 trillion lines of journal entry data — is the largest enterprise AI deployment in professional services history. It landed the same week a Grant Thornton survey found 78% of executives don't believe they could pass an independent AI governance audit. Both things are true simultaneously.
AI Agents Score Half as Well as PhDs on Real Work. Benchmarks Say Otherwise. Both Are Right.
Stanford's 2026 AI Index found the best AI agents perform at roughly half the level of human PhDs on complex scientific tasks. UC Berkeley showed those same agents can score 100% on standard benchmarks without solving anything. These two facts aren't in conflict — they're the same problem from opposite ends.
AI Agents Are Running Payroll Now. The Stakes Just Changed.
ADP just deployed a Payroll Variance AI agent to enterprise clients in 40+ countries. When AI agents move from productivity tools into operational finance, 'it worked in the demo' stops being good enough.
OpenAI Gave Agents a Sandbox. What They Still Need Is a Report Card.
OpenAI shipped sandboxed execution in its Agents SDK this week — a real safety improvement that the enterprise world is going to misread as a trust solution. Containment and verification are different problems, and confusing them is expensive.
96% of Enterprises Have AI Agents. Only 12% Know How to Govern Them.
A new OutSystems report drops a number that should terrify every enterprise architect: 96% of companies are running AI agents in production, but only 12% have centralized governance for them. Agent sprawl isn't a future problem — it's the problem you have right now.
The Liability Era for AI Agents Has Arrived
Gartner just told general counsels to buy AI insurance. The Register noted there's nobody to sue when agents fail. With enterprises averaging $207M in AI spend this year and Gartner projecting 2,000+ 'death by AI' legal claims by year-end, the question of who's liable when agents break things is no longer academic.
The Agent Marketplace Moment Is Here — And It's Moving Fast
monday.com just launched an AI agent hiring marketplace built with Anthropic. MCP hit 10,000 servers. A2A is now Linux Foundation infrastructure. The thesis SignalPot was built on is being validated in real time.