The AI news cycle doesn't slow down, but today's signal cuts through the noise: agents are burning budgets, CEOs are questioning everything, and even Amazon got burned by bad AI advice. Here's your daily AI pulse.
What's Breaking
One agent ran for 63 hours and burned $4,200
Production AI agents are entering infinite tool-call loops that drain budgets fast. One agent ran 63 hours straight, consuming $4,200. Another logged 319 retries on a single rate-limited task. Soft loops — where the agent rephrases slightly each time — are nearly invisible to monitoring. The fix isn't better prompting; it's infrastructure: circuit breakers, step-count limits, and hard budget gates. (Production postmortem)
56% of CEOs see zero ROI from AI spending
PwC found more than half of CEOs report neither increased revenue nor reduced costs from AI. Only 12% see both. Gartner confirms: 72% of organizations are breaking even or losing money. MIT found AI cheaper than humans in only 23% of vision tasks. Meanwhile, companies invest 93% of AI budgets in tech and just 7% in training people. (Computing)
Amazon's AI agent caused a production outage from outdated wiki data
Amazon's retail site went down in March after an engineer followed advice from an AI agent hallucinating from an outdated internal wiki. Even Amazon — one of the most sophisticated AI shops on the planet — got burned. SWE-Bench Pro shows top models resolve only 42-46% of realistic issues. No one's immune. (Wharton analysis)
Top 5 AI News
Anthropic's "Dreaming" — agents that learn from their own mistakes
Anthropic's biggest agent platform update lets agents review past sessions, curate memories, and self-improve. Outcomes evaluation and multi-agent orchestration hit public beta. Combined with a $50B raise at $900B valuation and rival Wall Street JVs (Anthropic's $1.5B with Blackstone/Goldman vs. OpenAI's $4B with TPG/Brookfield), the race for enterprise AI budgets is now being fought with forward-deployed engineers.
SAP acquires Prior Labs for €1B+, blocks unauthorized agents
SAP's €1B+ acquisition of German tabular AI startup Prior Labs comes with a power move: blocking all non-authorized agents, allowing only NemoClaw (Nvidia) and Joule Agents. The platform wars for AI agents are getting territorial.
Sierra raises $950M at $15B — enterprise agents get real
Bret Taylor's Sierra closed $950M at $15B with $150M ARR and 40%+ of the Fortune 50. They launched Ghostwriter, an agent-as-a-service product. Enterprise agents aren't hype anymore — the revenue says so.
Cerebras IPO: $3.5B at $26.6B — Nvidia's first real rival goes public
The largest tech IPO of 2026 so far. $10B in orders for a $3.5B offering at $26.6B valuation. As the first major Nvidia competitor to IPO, Cerebras is a bellwether for whether AI chip competition is real.
5,000 vibe-coded apps expose sensitive corporate data
RedAccess found 380K publicly accessible vibe-coded assets, 1.3% with sensitive data — roughly 5,000 exposed apps. The Lovable CVE affects 170+ production apps. Easy AI app-building created a massive attack surface most security teams don't know exists. (RedAccess)
Papers That Matter
The Impossibility Triangle of Long-Context Modeling — arXiv:2605.05066
Proves no architecture can simultaneously be efficient, compact, AND have good recall in long-context tasks. Classified 52 architectures — none escape the triangle. If you're betting on ever-longer context windows, this is your reality check. Read the paper →
EMO: Emergent Modularity in Mixture-of-Experts — Allen AI
Allen AI's model activates just 12.5% of experts for a specific task with near-full performance — standard MoE models degrade severely at partial usage. This opens the door to composable AI: deploy only the capabilities you need, when you need them, without retraining.
What This Means For You
Today's pain points share a common thread: the gap between what AI can do and what it reliably does in production. That $4,200 agent loop didn't happen because the model was bad — it happened because nothing stopped it. The 56% of CEOs seeing no ROI aren't hitting a technology wall; they're hitting an implementation wall. Amazon's outage wasn't a model failure — it was a grounding failure.
Most organizations deploy agents with the operational maturity of a prototype. Circuit breakers, budget caps, source verification — these aren't optional. They're the difference between an AI investment that compounds and one that hemorrhages money. If your monitoring can't detect 300 retries of the same failed query, you're flying blind.
The market is bifurcating. SAP builds walled gardens; Anthropic bets on self-correcting agents. Both attack the same reliability problem. The companies that sort out governance before their first expensive weekend will be the ones that see actual ROI.
Written by The AI Architect team at Atobotz