“tokenmaxxing” Is Making Developers Less Productive Than They Think

There’s an aged saw successful management: What you measurement matters. And, typically, you get much of immoderate you’re measuring.

Software engineers person debated productivity metrics for decades, starting pinch lines of code. But arsenic nan caller procreation of AI coding agents delivers much codification than ever, what their managers ought to beryllium measuring is little clear.

Enormous token budgets—essentially, nan magnitude of AI processing powerfulness a developer is authorized to consume—have go a badge of grant among Silicon Valley developers, but that’s a very weird measurement to deliberation astir productivity. Measuring an input to nan process makes small consciousness erstwhile you presumably attraction much astir nan output. It mightiness make consciousness if you’re trying to promote much AI take (or trading tokens), but not if you’re trying to go much efficient.

Consider nan grounds from a caller people of companies operating successful nan “developer productivity insight” space. They’re uncovering that developers utilizing devices for illustration Claude Code, Cursor, and Codex make a batch much accepted codification than they did before. But they besides find that engineers person to return to revise that accepted codification acold much often than before, undercutting claims of accrued productivity.

Alex Circei, nan CEO and laminitis of Waydev, is building an intelligence furniture to way these dynamics; his patient useful pinch 50 different customers that employment much than 10,000 package engineers. (Circei has contributed to TechCrunch successful nan past, but this newsman had ne'er met him before.)

He says that engineering managers are seeing codification acceptance rates of 80% to 90%—meaning nan stock of AI-generated codification that developers o.k. and keep—but they’re missing nan churn that happens erstwhile engineers person to revise that codification successful nan pursuing weeks, which drives nan real-world acceptance complaint down betwixt 10% and 30% of generated code.

The emergence of AI coding devices led Waydev, founded successful 2017 to supply developer analytics, to wholly rework its level successful nan past six months to reside nan proliferation of accelerated coding tools. Now, nan institution is releasing caller devices that way nan metadata generated by AI agents, offering analytics connected nan value and costs of their codification to supply engineering managers pinch much penetration into some AI take and efficacy.

While analytics companies person an inducement to item nan problems they find, nan grounds is mounting that ample organizations are still figuring retired really to usage AI devices efficiently. Major companies are noticing—Atlassian acquired DX, different engineering intelligence startup, for $1 cardinal past year, to thief its customers understand nan return connected finance connected coding agents.

The information from crossed nan manufacture tells a accordant story: More codification is being written, but a disproportionate magnitude of it isn’t sticking.

GitClear, different institution successful this space, published a report successful January that recovered AI devices accrued productivity, but besides that its information showed “regular AI users averaged 9.4x higher codification churn than their non-AI counterparts”—more than double nan productivity gains nan devices provided.

Faros AI, an engineering analytics platform, drew connected 2 years of customer information for its March 2026 report. The finding: codification churn—lines of codification deleted versus lines added—had accrued 861% nether precocious AI adoption.

Jellyfish, which bills itself arsenic an intelligence level for AI-integrated engineering, collected data connected 7,548 engineers successful nan first 4th of 2026. The patient recovered that nan engineers pinch nan largest token budgets produced nan astir propulsion requests (proposed changes to a shared codebase), but nan productivity betterment didn’t scale. They achieved 2 times nan throughput astatine 10 times nan costs of tokens. In different words, nan devices are generating volume, not value.

These kinds of statistic ringing existent erstwhile you talk to developers, who are uncovering that codification reappraisal and method indebtedness are stacking upm moreover arsenic they revel successful nan state of nan caller tools. One communal uncovering is nan quality betwixt elder and inferior engineers, pinch nan second accepting acold much AI-generated code, and dealing pinch a larger magnitude of rewriting arsenic a consequence.

Still, moreover arsenic developers activity to understand precisely what their agents are up to, they don’t expect turning backmost anytime soon.

“This is simply a caller era of package development, and you person to adapt, and you are forced to accommodate arsenic a company,” Circei told TechCrunch. “It’s not for illustration it will beryllium a rhythm that will pass.”