What three rows does the engineering-manager scorecard need after May 22?

I sat with an engineering leader last week who pulled up the one-on-one template she had been using since January. Seven rows. Delivery against roadmap. People management. Cross-team alignment. Hiring pipeline. Career conversations. Skip-level notes. Risks. We read through it together. By the time we finished, four things had happened in the news that her template did not cover.

This is what I want to walk through. Because the engineering-manager scorecard most leaders are running right now was written for a team made of humans, and the team being managed in May 2026 stopped being that team. The June calibration meeting is going to expose every leader still operating on the January template.

What changed in 48 hours

The week of May 22-23 carried four things landing inside the same window.

Fortune published a piece by Jake Angelo on May 22 with a verbatim quote from Bryan Catanzaro, the Nvidia VP of Applied Deep Learning. He said, “For my team, the cost of compute is far beyond the costs of the employees.” The same article confirmed Microsoft cancelled most of its internal Claude Code licenses six months after encouraging adoption, and quoted Uber’s CTO Praveen Neppalli Naga, who told The Information in April that Uber had burned its entire 2026 AI coding tools budget in four months.

CFO Dive published Grace Noto’s reporting the same day on Intuit’s 17 percent workforce reduction. About 3,000 employees. Restructuring charges between $300 and $340 million. CFO Sandeep Aujla framed the structural logic in his own words: a leaner structure that “will accelerate how we operate with greater focus, speed, agility and an even stronger commitment to profitability.” CEO Sasan Goodarzi added the cultural framing about staying “a company of builders.”

Anthropic shipped Claude Code v2.1.147 through v2.1.150 across May 22 and 23. The release notes name what changed precisely.

Four named categories. Skills. Subagents. Plugins. MCP servers. Each one is a row that did not exist in the harness output a week ago.

Salesforce Ben’s Sasha Semjonova covered Marc Benioff’s Agentforce Coworker on the same May 22. Benioff’s verbatim framing: “Your new AI teammate is now inside every Salesforce search bar. It doesn’t just find things, it works with you.” The Coworker reaches across CRM data, workflows, customer history, opportunities, and cases, then “takes action in real-time” inside Salesforce, Slack, Teams, ChatGPT, or wherever else people work most.

Four stories. None of them is a coding-tools announcement on its own. Together they describe the engineering-manager job in May 2026.

What most scorecards are still doing

The default engineering-manager scorecard I see in companies between Series B and pre-IPO has roughly the same seven rows I read with that leader. Delivery, people, alignment, hiring, careers, skips, risks. The rows have not changed since the era when a team meant a roster of humans with editors open and a payroll spreadsheet for the cost line.

That scorecard still works for most of what an engineering manager does. The problem is that it stops working the moment any of the following becomes true on the team: there is a daily token spend at the team level, there are persistent or scheduled agents executing work between standups, there is a cross-functional AI coworker reading the team’s APIs without a PM in the loop, or the company has named “leaner structure” as the planning posture.

By May 22 all four are true at the median Series B and above engineering org I talk to.

The three rows the scorecard is missing

Here is what I would add. Three rows, each grounded in something that landed this week.

Row one: token spend per direct report, per week. Catanzaro’s quote is not the average engineering team’s reality yet. Nvidia is unusually compute-heavy. But the cost ratio is moving in one direction, and the supporting evidence is named in the same article. Microsoft cancelled most of its internal Claude Code licenses. Uber’s CTO publicly said the 2026 AI coding tools budget was gone in four months. Goldman Sachs is forecasting a 24-fold token consumption increase by 2030, reaching 120 quadrillion tokens monthly.

What this means for a one-on-one slide: the team’s monthly compute number now needs a per-person line, in the same way the people-management row has a per-person line. Not because the EM is going to penalize the engineer who uses the harness most. Because the CFO is going to ask the CTO, the CTO is going to ask the EM, and the EM who walks in with a single roll-up number gets sent back to the keyboard. The CTO’s June board slide already has a row for compute. The EM’s one-on-one slide should reconcile to it.

Row two: per-subagent, per-skill, per-MCP utilization. This is the row Anthropic just made cheap. Before this week, putting subagent-level utilization on an EM scorecard meant building a custom dashboard or running pipeline jobs against OTEL exports. As of v2.1.149 the breakdown ships in the harness itself.

The slide row reads something like: top three skills by usage, top three subagents by cost, MCP servers consuming more than a defined floor. The EM does not need to know how the underlying token math works. The EM needs to know which subagents the team built that nobody is using anymore, which skill is silently driving most of the spend, and which MCP server is being invoked by code paths nobody named in the design doc.

Row three: cross-team AI coworker consumption. This is the row the Salesforce news made urgent. When Benioff says the AI teammate is in every search bar, what he is describing for the engineering team is that customer success, sales operations, and finance now have an AI coworker that will start invoking the engineering team’s APIs without asking. Not through a Jira ticket. Through a tool call.

The slide row used to read “internal NPS from PMs and CS.” The sibling row now reads “cross-team AI coworker traffic to my team’s interfaces, and the top three consuming workflows.” The EM whose API is suddenly the default tool for the Coworker has more on the line than the EM whose API never gets called. Both need to know which one they are.

What this pattern actually is

I have been describing this as a scorecard rewrite, but the underlying mechanism is simpler. The unit an engineering manager is managing flipped. It used to be a roster of humans with a payroll cost. By May 2026 it is a hybrid: humans with a payroll cost, plus persistent and ephemeral agents with a token cost, plus a set of APIs that other functions’ AI coworkers read on a schedule.

Each of the four pieces of news from May 22-23 confirms one face of this hybrid. Fortune confirms the cost line. Anthropic ships the per-agent observability. Salesforce makes the cross-function consumption real. Intuit and the parallel Meta storyline (Meta tracked engineers across Gmail, Chat, VS Code and internal tools to train internal models, then laid off 8,000 employees, per the BuildFastWithAI digest on May 23) confirm that companies have already named “leaner structure” as the planning posture, which puts the EM in the position of writing the why-this-team-still-exists paragraph before procurement asks for it.

There is also a Dell Tech World line I am going to steal. On SiliconANGLE’s coverage of the same May 22, a speaker said from the stage, “Put your agent in a sandbox of some kind and then give it read-only access and start feeding it with that access [level] and see what it can uncover.” That is the cheapest operational discipline an EM can write into a pre-merge checklist this quarter. Eighty percent of ClawCon attendees that day were first-time software builders, which tells me the discipline is going to be named as the industry default by the people learning to build with agents this year. The engineering managers who write it into the team runbook before that point look prepared. The ones who add it after the first incident look reactive.

What I’d tell you over coffee

Three moves before June calibration.

First, copy the existing one-on-one template, add the cost row and the utilization row, and pull the data from the /usage per-category breakdown plus the monthly tenant export from finance. The data exists now. The slide does not. Build the slide once, reuse it for the rest of the year.

Second, write the restructure-defense paragraph for the team before someone in procurement writes it for the team. Not the layoff paragraph. The why-this-team-still-exists paragraph. Name the work that requires the team’s specific judgment, the cross-team AI coworker dependencies that route through the team’s interfaces, and the agent-supervision capacity the team has built that another team cannot copy in a quarter. This paragraph is now an EM deliverable, not a CTO deliverable. The leader who walks into June calibration with it already written changes the meeting.

Third, write the sandbox-first rule into the pre-merge checklist for any agent change that touches a credential, a production system, or a customer-facing surface. Read-only first. Expand the surface only after observed behavior. It is the cheapest single line a leader can add to the team’s runbook, and it is the one that keeps a team off the post-incident list when the harness of the week ships an unexpected capability.

The scorecard rewrite is not optional. The harnesses just shipped the data. The CFO already has the narrative. The cross-function AI coworkers are already reading the APIs. The only choice is whether the rows show up on the EM’s slide in June, or whether the CTO has to add them on the way to the boardroom in July.