Every week I read AlphaSignal, The Batch, Exponential View, Tunguz, The Rundown, and about ten more AI newsletters. Most of them cover the same stories. This is where I pull the signal from the noise and write what actually matters for people building production systems.
This week had one theme: the governance deficit. Capability is accelerating. Controls are not. Every major story this week, from code quality data to enterprise surveys to frontier model launches, pointed at the same gap.
I spent part of this week reviewing how teams are actually deploying agents in production. The pattern is consistent. The capability is there. The guardrails aren’t. And the gap is widening faster than most orgs are closing it.
Here’s what you need to know.
THE BIG FOUR
1. 22,000 Developers Measured. AI Ships Faster. It Also Breaks More.
Faros.ai published its 2026 Engineering Report covering 22,000 developers across 4,000+ teams over two years.
AI code acceptance rose from 20% to 60%. Epics completed per developer climbed 66%. Teams are producing more, faster.
But production incidents per pull request surged 243%. Code churn (lines deleted vs. lines added) increased 861%. And 31% more PRs are merging with no human review at all.
This is the first large-scale measurement of what happens when AI-assisted coding reaches majority adoption. Velocity increased. Quality decreased. Shipping 66% more epics means nothing if incident rates triple.
2. Five Surveys Converge: Enterprises Deploy Agents Without Controls
96% of IT leaders use agents in workflows (~1,900 respondents, OutSystems). Only 12% have a centralized platform to control them. 78% would fail an AI governance audit within 90 days (~950 leaders, Grant Thornton). Only 20% have tested an incident response plan for AI failures.
95% of C-suite leaders have an AI strategy with an average commitment of $186M (2,110 leaders, KPMG). Only 9% orchestrate multi-agent workflows. 62% of enterprises experiment with agentic AI while only 23% run it in production (McKinsey data, cited at RSA Conference 2026). 29% of employees deploy shadow agents without IT oversight (Microsoft Cyber Pulse).
The standout number: companies using governance tools deploy 12x more AI projects to production (20,000+ organizations, Databricks, covering 60% of Fortune 500).
That 12x is correlation, not proven causation. But the signal is strong. Governance is the mechanism that gets AI from pilot to production. Spending $186M on AI strategy while 78% would fail an audit is funding capability without the infrastructure to deploy it. That’s the governance deficit.
3. Three Frontier Coding Agents in One Week
Anthropic released Opus 4.7 (87.6% SWE-bench Verified, 64.3% SWE-bench Pro). OpenAI expanded Codex into a desktop superapp with parallel agents and session memory (3M weekly users, 70% month-over-month growth). Google’s Gemini 3.1 Pro leads ARC-AGI-2 at 77.1%.
Three frontier coding agents. Each more autonomous than its predecessor. More autonomy means more unsupervised decisions, more API calls, more code changes, more production deployments without the governance infrastructure the surveys described. The coding agent war accelerates the need for controls.
4. Snap Cuts 16%. Cloudflare Cuts 99.9% of Tokens. Agent Ops Is a Discipline.
Snap cut 1,000 jobs (16% of its 5,261 workforce) citing AI that now writes 65% of new code and handles over a million monthly queries. The company targets $500M in annual savings by end of 2026.
Cloudflare shipped an MCP server that reduced token consumption from 1.17M to roughly 1,000 tokens across 2,500+ API endpoints. 99.9% reduction. The difference between a useful agent integration and one that burns through budgets is often a single architectural decision about how tools describe themselves to models.
Snap shows what happens when agent capability works and organizations restructure around it. Cloudflare shows what happens when engineering teams optimize the agent-to-tool interface. Without either discipline, you get runaway loops, unpredictable costs, and zero useful output. Agent operations (observability, circuit breakers, cost ceilings, deployment governance) is an emerging discipline that most enterprises haven’t built yet.
ALSO WORTH KNOWING
EU AI Act high-risk obligations proposed to slip to December 2027. On March 26 the European Parliament voted 569-45-23 to delay Annex III high-risk obligations from August 2, 2026 to December 2, 2027. Watermarking rules for AI-generated content move to November 2, 2026. The Council still has to ratify, and legal advisors recommend planning as if the August 2026 deadline holds until that vote lands. Fines remain at 30M EUR or 6% of global turnover. Only 3 of 27 EU member states have designated AI authorities. European AI spending is projected at $290B by 2029. The runway widened. The governance deficit did not.
Google Gemma 4: open models for local agents. The family spans 2B to 27B parameters with native function calling and 256K context. The 26B MoE variant runs only 3.8B active parameters per token, achieving 100+ tokens per second on a MacBook Pro M5 Max. Day-zero MLX and vLLM integration. Open-weight local agents are becoming production-viable for enterprises that need on-premises control.
GPU prices climbed 48% in two months. Blackwell chip cloud rental hit $4.08 per hour, up from $2.75 (Ornn Compute Price Index). CoreWeave raised prices 20% and extended minimum contracts from one year to three for smaller customers. Compute scarcity is real. Meanwhile, Allbirds pivoted to GPU-as-a-service and its stock jumped over 600%.
Bank compressed 6-month orchestration rebuild to 6 days./span> An SD Times case study documented a top-10 bank that collapsed a six-month agent orchestration rebuild into six days using next-generation tooling. The gap between what’s possible and what’s deployed keeps widening.
THE PATTERN
This week gave the governance deficit numbers.
- • Faros measured it in code quality: 60% AI acceptance, 243% more incidents. Five surveys measured it in deployment: 96% run agents, 12% control them. Databricks measured it in outcomes: 12x more production deployments for companies with governance tools.
- • Then three frontier coding agents launched in seven days. Each one more autonomous. Each one amplifying the gap.
- • The specific scaffold that matters right now is operational governance. Who approved this agent. What can it access. How much can it spend. What happens when it fails.
Every number this week points to the same bottleneck. The 78% who would fail an audit just got 16 more months of runway, if the Council ratifies the delay. Most won’t use it.
I’ve been building production systems for 25 years. The pattern is always the same: the tool arrives before the discipline to use it safely. The firms that close the accountability gap first don’t just survive the transition. They define the standard everyone else follows.
Sources: Faros.ai, OutSystems, Grant Thornton, KPMG, Databricks, McKinsey (via SoftServe RSA 2026), Microsoft Cyber Pulse, Anthropic, OpenAI, Google, Gartner, Forrester, The Rundown AI, AlphaSignal, Tomasz Tunguz/Theory Ventures, Cloudflare, InfoQ, SD Times, Dyno Therapeutics, IDC, EU AI Act, Ornn Compute Price Index
I write about Production AI, enterprise AI adoption, and building systems that actually work. Follow along if that’s your thing.
Latest Articles
Discover materials from our experts, covering extensive topics including next-gen technologies, data analytics, automation processes, and more.
