Mr. Irrelevant

93% — the rate at which Claude Code users approve permission prompts without inspecting them, per Anthropic's own published telemetry this week. Two out of three "human in the loop" approvals are not approvals. They are a checkbox the human has learned to click. Anthropic put a number on the failure mode their entire user base relies on for safety — and shipped the alternative the same week. The encyclical the Pope released Monday called for vigilance. The data Anthropic released Monday says nobody is doing it. Mr. Irrelevant works in your building too.

Stop ranking employees by token consumption. Start measuring useful output per token. The token dashboard is a 40-yard dash. It measures input, not output, and it punishes the engineer whose work-quality-per-token is highest. Uber gamified its Claude Code spend into a leaderboard and blew its 2026 AI budget by April. Microsoft cancelled Claude Code licenses for ~100K engineers in response. The Fortune 50 power user on Maxinomics' viral X post last week had her tokens rationed for being too productive. The dashboard that produced those decisions is the same dashboard most companies are running today. Replace it before the next round of cuts.

Read Anthropic's containment essay and the 28-integration announcement with your CTO, CISO, and audit-committee chair on the same call this week. Then answer one question on the same page: what does our equivalent framework look like? The labs publish doctrines now. Anthropic just defined the genre. The companies that publish a containment framework by Q4 will have a moat against the ones that haven't thought it through. The companies that don't will look in 2027 like the ones that didn't have a GDPR policy in 2017. The 28 integrations include CrowdStrike, Palo Alto, Okta, Datadog, Cloudflare, IBM, Microsoft, and twenty-one others — the procurement path is already built.

Hire your first orchestrator. This is the role that doesn't exist on your current org chart and will exist on the 2027 org chart. The orchestrator is a senior IC whose direct reports are agents, not humans. They tune the fleet, write the meta-prompts, build the eval harnesses, own the output. Their seat at the senior-staff table should match the VP whose human-team is now smaller than the orchestrator's fleet. The headcount-prominence org chart is over. Agents of note reporting to you is the new flex. Start with the role title and the comp band before someone else hires the candidate you didn't know you were looking for.

Keeping up with AI is hard. We know — we do this daily. Today's actions touched on token-measurement redesign, AI-safety doctrine, and the orchestrator role — all work we've done with clients this quarter. If your team is staring at "redesign our AI Combine" and isn't sure where to start, that's the conversation we're built for.

Outsider Labs Advisory →

Wednesday, May 27

1. Anthropic shipped the AI-safety stack the day after the Pope. Twenty-eight enterprise security integrations (CrowdStrike, Palo Alto, Okta, IBM, Cloudflare, twenty-three others), a Compliance API that pipes conversation and event logs into existing SOC dashboards, a published containment essay naming Claude Mythos as a model they held back because the blast radius was too high. The other frontier labs do not have a comparable framework on the public record. That is what category capture looks like. (Full analysis below.)

2. 93% of Claude Code permission prompts get rubber-stamped. That number comes from Anthropic's own telemetry, published this week. It is the cocktail-party grenade of the AI moment: the "human in the loop" is checkbox theater, and the lab that built the loop just published the proof. Every CISO claiming permission prompts are sufficient safety needs to read the essay this week and revisit the deployment doc next week. (Full analysis below.)

3. China just embargoed AI talent. Alibaba and DeepSeek researchers now need state approval for overseas travel, per techmeme coverage this week. After chips (2022) and energy/capex (2024), talent is the third leg of AI export control. The global research market just got rebalkanized — half the candidate pool for frontier roles in 2027 is going to disappear by passport color.

4. Grok Build CLI launches the week Microsoft cancels Claude Code. TLDR confirmed the xAI coding agent dropped in beta for SuperGrok and X Premium Plus this week — at $8/month. With DeepSeek V4-Pro cutting prices 75% the same week, the coding-agent market just stratified into five distinct tiers. Picking yours is now as politically loaded as picking your CRM in 2018, and the lock-in is faster.

5. The silicon layer repositions for agents. Three concurrent stories on the same Tuesday: NVIDIA's Vera CPU (88 custom Olympus cores, 1.2 TB/s memory bandwidth) gets a Phoronix benchmark calling it the most formidable x86_64 competitor ever. Qualcomm strikes a millions-of-ASICs deal with ByteDance for the Doubao agents. Meta open-sources a Blackwell attention kernel claiming a 2.3x speedup. Agents don't just want more GPUs anymore. They want a different motherboard.

The Combine Is Yours

Anthropic published two artifacts this week that almost no other newsletter is going to put on the same page. Monday, the engineering blog ran How We Contain Claude Across Products — an essay that named, in plain language, the rate at which their own users approve permission prompts without inspection: ninety-three percent. Two out of three "human-in-the-loop" approvals are theater. The same essay disclosed that Claude Mythos Preview, a next-generation agent the lab had built, was held back from release in April 2026 because its blast radius was deemed too high. The model exists. The model is not shipping. The containment around it was not ready.

The other artifact landed the same week: twenty-eight enterprise security integrations, a Compliance API that streams conversation content and activity logs into the same dashboards CrowdStrike, Palo Alto, Okta, Datadog, Cloudflare, and IBM customers already monitor. The list reads like a procurement bill of materials for an enterprise SOC. That is the point. Anthropic is not asking the SOC to learn how to monitor an agent. They are shipping the agent into the monitoring infrastructure the SOC already runs. The other frontier labs — OpenAI, DeepMind, xAI — do not have a comparable framework on the public record. The category is, this week, a one-lab category.

Two days ago Pope Leo XIV released Magnifica Humanitas. The encyclical's central claim, the one CEOs should be cross-stitching onto a sampler, is seven words: AI is a tool that requires vigilance. The vigilance is not engineering. The vigilance is corporate. It is the work of governing what your organization does with the agent — who has access, to what, under what audit, with what fail-stop. That is the layer above what Anthropic just shipped. Above that sits the engineering question of whether the individual agent is safe to deploy at all — the layer Anthropic's essay addressed and the layer where the lab kept Mythos in the box.

Three layers, three owners. Agent-level safety belongs to your engineering team. Corporate-level safety belongs to your CIO and your board. Technology-level safety belongs to the lab — and as of this week, the lab has shipped a product. You can buy it. The other two layers are yours.

Here is where the org chart breaks. For forty years, prominence inside a Fortune 500 was measured in humans-who-report-to-you. The director had three. The VP had twenty-eight. Compensation, title, seating chart, all of it pegged to the count. The hardest thing in a 1985 corporation was getting twenty humans aligned on a single goal — and the senior director who could do that deserved the senior director's pay. The unit of leverage moved. The hardest thing in 2026 is getting twenty agents aligned, and figuring out which two humans in the building can do it. The number of humans reporting to that orchestrator is not the indicator. It might be zero. The number that matters is the quality of the agent fleet they orchestrate and the output it generates. Agents of note is the new flex. Your org chart does not yet have a row for them.

The same week corporate America is being told the agent-orchestration role is the new senior-VP, the dashboards most enterprises are using to find their orchestrators are still measuring tokens consumed. The token leaderboard is a 40-yard dash. It is run on the engineer, the lawyer, the customer-success rep, the FP&A analyst — five completely different jobs ranked on one universal input metric. The 40-yard dash never told you who could read a defense. The token dashboard never tells you who can diagnose what an agent did wrong and write a better prompt before lunch. Different game. Same broken stopwatch.

The bill, when it lands, will look exactly like the hundred-thousand technology-sector layoffs already cut in the first five months of this year — except the next hundred thousand will be cut on metrics no one in the building can yet name. The companies that win the decade are the ones whose front office is rebuilding the Combine this quarter. The labs are not going to do that for you. The Pope is not going to do it for you. The consulting firm sells the old Combine — that is, structurally, their business model. The redesign is yours, in your building, on Tuesday.

Don't run the Combine. Redesign it.

At COAI today: Full essay on the three-layer AI-forward operating system, the agents-of-note org chart pivot, and what corporate America is missing in its current AI Combine at getcoai.com/news-letter/mr-irrelevant/.

— Harry and Anthony

Sources: