Breaking

Anthropic Engineer Says Over 90% of Its Developers Now Build With 'Self-Improving Loops'

July 20, 2026 at 16:45 EDT

AI Agents
Software Dev & Coding
Foundation Models

An Anthropic engineer says more than 90% of the company's engineers are now building software with "self-improving loops," and predicts that figure will reach 100% within four to six months. The remark, made during a recent podcast appearance, has become a rallying point for a broader shift in how AI-native teams write code.

May 2026 · Anthropic

From prompts to loops: inside Anthropic's self-improving code machine

More than 90% of Anthropic's engineers now build software with "self-improving loops" — autonomous agents that set goals, execute, verify and iterate. An engineer predicts that figure reaches 100% within four to six months.

90%+

of engineers already building with self-improving loops

80%+

of merged code authored by Claude as of May 2026

76%

task success on open-ended benchmarks — up ~50pts in 6 months

Code merged per engineer, per day

Median lines merged in Q2 2026 ran roughly 8× the 2024 figure.

1×

2024

8×

Q2 2026

"Loop engineering" — the recursive workflow

Instead of prompting step by step, engineers design autonomous loops that run for days.

Set goal

→

Execute

→

Verify

→

Iterate ↺

The believers

A March 2026 survey of ~130 employees estimated median output ~4× higher with AI assistance. Engineers now review and direct while Claude writes — and advocates urge teams to "design loops, not prompts."

The skeptics

Critics say the hard part is building reliable verifiers to check an agent's output. Running agents continuously for days raises cost and reliability concerns — and the internal metrics can't be independently verified.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.