The harness productivity number your board will ask about after reading the Coinbase letter

Two things happened on Wednesday that a board director will quietly read together by Friday.

The first was Code with Claude 2026 in San Francisco. Anthropic walked through Claude Managed Agents, a multi-agent orchestration layer, and a new async-automation feature called Routines. Simon Willison’s live notes captured the headline statistic: API volume is up 17x year-on-year on the Anthropic platform. Mercado Libre, with 23,000 engineers, was named onstage as targeting 90 percent autonomous coding by Q3 of this year. Shopify, ServiceNow, Accenture, and Netflix all got the customer-logo treatment.

The second was Coinbase. Brian Armstrong sent a 7am email cutting roughly 14 percent of the workforce, around 700 people. The line that traveled was his: “I’ve watched engineers use AI to ship in days what used to take a team weeks.” IBTimes UK and TechTimes both quoted it verbatim within hours.

Same Wednesday. Two very different rooms. One number not to put in a board pre-read. One conversation worth having anyway.

What every CEO has been quietly running the math on this week

If a public company can attribute 14 percent of its workforce reduction to AI engineering productivity, and a vendor can put a 23,000-engineer customer on stage promising 90 percent autonomy by Q3, what’s the productivity number for the rest of us? And if there isn’t one, are we underinvesting?

Three answers I’m seeing this week, all of them flawed.

Copy Coinbase. “We’re seeing the same thing internally.” That gets a follow-up question a CFO will not enjoy. What percentage of engineering output is verifiably attributable to coding agents this quarter? Most companies have not instrumented this. The post-hoc attribution is hand-waving, and a sophisticated board director hears the hand-wave inside ten seconds.

Copy Mercado Libre. “We’re targeting 90 percent autonomous coding by end of year.” This is a Q3 forward commitment from a 23,000-engineer engineering org with a deep platform team. Most Series B and Series C companies do not have the platform capacity to enforce 90 percent autonomy without a verification stack that costs more than the productivity gain it produces. A target a company cannot operationalize is worse than no number, because it gets re-quoted in pre-IPO conversations.

Cite Anthropic. “API volume is up 17x year-on-year.” Lovely. That number tells the board how Anthropic’s business is doing. It tells them very little about a specific engineering org.

I keep watching the same pattern. An executive grabs the most public number on offer that week, the board nods, and someone six months later asks what actually shipped.

Where the believer and the skeptic on a board diverge

There is a productivity-believer and a productivity-skeptic on every board I have ever worked with. They are each waiting to hear something different.

The believer wants a number that goes up and to the right. The skeptic wants a number that ties to a P&L line. The trap is that vendor numbers (17x API volume, 4 percent of public commits, 90 percent autonomous targets) satisfy the believer and infuriate the skeptic. Coinbase-style attribution numbers infuriate the believer because they sound rhetorical, and they make the skeptic suspicious because the same week’s Coinbase earnings showed retail transaction revenue down 45 percent and Q1 revenue guided down 26 percent year-over-year. AI productivity is the public framing. Crypto cycle pressure is the math underneath.

That does not make Armstrong wrong. Engineers shipping in days what used to take weeks is real for some teams, on some workloads, on some weeks. It is also a quote that costs nothing to put in a layoff letter, and that is exactly why a sharp director wants three things underneath any number presented to them.

A source that is not the vendor.

A measurement window the CFO can verify.

A counterfactual. If this productivity is real, what does it change about the hiring plan, the gross margin, or the roadmap, this quarter, in dollars?

If a number does not survive those three filters, it is decoration. The Code with Claude keynote and the Coinbase letter both had decoration in them. The job is to pull out the part of each that survives.

The three numbers an executive can actually own this quarter

Engineer-month-equivalent of coding-agent work shipped to production, this quarter. Pull it from the CI/CD merge log, not the harness vendor’s dashboard. Filter to commits where at least one PR was authored by an agent. Count the engineer-months at the loaded cost from finance. This is the gross productivity contribution number, in dollars, defensible to a CFO because the numerator and the denominator both came from inside the company.

Verified-output rate. Of the agent-authored PRs that merged this quarter, what fraction passed without a reviewer-induced rewrite within seven days? This is the quality discount on number one. If the verified-output rate is below 70 percent, the gross productivity number is mostly fiction. Mercado Libre’s 90 percent autonomous coding target only matters if their verified-output rate stays high. Without that, 90 percent autonomous is 90 percent rework.

Headcount-plan delta attributed. This is the only one that touches the 2027 hiring plan, which is the conversation a board director actually wants. Take the Q3 hiring plan from December. Adjust it for productivity contribution and verified-output. Show the delta in dollars, and own the number with a name (the executive’s own), a date (today), and a re-evaluation review (next quarter). If the delta is zero, that is also a valid answer, and it is more credible than a fabricated 14 percent.

These three numbers, in this order, are what I would put in front of my own board this week. None of them came from Code with Claude. None of them came from the Coinbase letter. All of them came from data the company already has on disk.

If a number does not survive a vendor-source filter, a CFO-verification filter, and a dollars-this-quarter filter, it is decoration.

What I would tell you over coffee

The Coinbase letter is not a productivity benchmark. It is a public-company statement about a cost-cutting decision that needed a narrative, layered on top of a weak crypto quarter. The Code with Claude keynote is not a productivity benchmark either. It is Anthropic doing its job, which is showing customers that the product gets better. Both are useful. Neither belongs in your board deck without three numbers of your own underneath them.

The good news. The data is almost certainly already in the building. CI/CD logs, harness usage exports, finance’s loaded cost per engineer. That is a one-week project for one person, and it changes the productivity conversation in the board pre-read from rhetorical to specific.

That is enough to walk in calm.