A counterintuitive math problem: if you have an Agent that's 95% accurate at each step -- sounds pretty good, right? -- what's the probability of a correct final result after it executes 20 consecutive steps?

The answer is 35.8%.

That's the "17x error amplification effect" discovered in joint research by Anthropic and Material (500+ tech leaders surveyed). 0.95 to the 20th power is approximately 0.358; the error rate amplifies from 5% per step to 64% end-to-end. And that's the optimistic estimate -- in reality, Agent per-step reliability often doesn't even reach 95%.

Three Bloody Real Incidents

The research disclosed three chilling real cases.

The first was an AI sales Agent's "discount gate." A SaaS company deployed an AI sales Agent to automatically handle customer renewal negotiations; the Agent was granted some discount authority. But because capability and permission weren't strictly separated -- the Agent had the capability to give 50% discounts, and was indeed configured with that permission -- it unilaterally offered a 50% discount in a major client negotiation without requesting anyone's approval. By the time the sales director found out, the contract had already been sent.

The second was a connector project's $500,000 tuition. Five senior engineers spent three months building enterprise system connectors to let Agents interface with clients' SAP and Oracle systems. The project completely failed, losing over $500,000. The failure reason wasn't technical impossibility, but severe underestimation of system integration complexity -- 46% of surveyed teams listed "system integration" as the #1 obstacle to Agent deployment.

The third was brute-force RAG's "information overload crash." A customer service Agent was fed too many documents as context; the output quality collapsed due to information overload -- not wrong answers per se, but it started talking nonsense, applying Product A's after-sales policy to Product B.

Implementation Recommendations: P0/P1/P2 Priority Checklist

The report gives tiered practical recommendations, ordered by priority:

P0 (Must do): Minimum-permission configuration; Agents default to zero permissions, each permission must be explicitly granted; sandbox rehearsal -- any write operation by an Agent to production must first be validated in sandbox; human approval checkpoints -- operations involving money, data deletion, or permission changes must have human confirmation.

P1 (Should do): Step count control -- single Agent task chains shouldn't exceed 5 steps; complex tasks split into multi-Agent collaboration; error retry mechanisms -- critical steps with retry and rollback; complete operation logs -- every Agent action must be traceable.

P2 (Recommended): Gradual canary rollouts starting from 1% traffic; cost budget limits -- set Token consumption caps for Agents; A/B testing frameworks -- new and old flows running in parallel for comparison.

Core Cognitive Shift

The most important value of this research isn't the technical recommendations, but revealing a cognitive mistake: many teams treat Agents as "super-employees" -- smart, reliable, tireless, safe to hand tasks to. But reality is that Agents are more like "very capable interns with no common sense" -- they can do many things, but you must set clear boundaries, strict processes, and brakes that can stop them at any time.

86% of teams already use Agents, but only 17% have achieved true scaled deployment. This gap isn't a model capability gap; it's an engineering and governance capability gap. In H2 2026, Agent competition will shift from "who can make a Demo" to "who can run Demos safely and stably in production."

See you tomorrow.