Anthropic writes
80% of its code with AI.
What that means for the mid-market.

In May 2026, Anthropic wrote more than 80 percent of its new production code with its own model, Claude. Not in a test environment. In production. Vercel automates 96 percent of its marketing drafts and 93 percent of its support. Cognition insures its AI productivity for up to ten million US dollars. That sounds like Silicon Valley. But we build for others what we use ourselves — and we see firsthand: the same lever works in the German mid-market too.

What happens when AI writes its own code?

Anthropic calls its research report "When AI Builds Itself." The title is meant literally. In February 2025, the share of AI-written code internally was still in the low single-digit percent range. Fifteen months later: more than 80 percent. That's an eightfold increase in code volume per developer per quarter compared to the 2021–2025 baseline.

Human developers continue to act as reviewers and architects. They decide what gets built. Claude takes over the how. The success rate on complex, open-ended engineering problems rose to 76 percent over the same period — a jump of 50 percentage points in six months.

"Claude-written code was somewhat worse than human-written code at Anthropic in late 2025, is roughly at parity today, and we expect it to be strictly better within the year."

This isn't hype. It's a forecast backed by data. Günter Born summarized the report on Borncity on June 7. VentureBeat documented the original data on June 4. The numbers hold up. What they mean for your company depends on whether you read them as a headline — or as an invitation.

Three companies leading the way

Cognition: Devin Desktop and the ten-million-dollar guarantee

On June 7, 2026, Cognition unveiled its new flagship: Devin Desktop. The software combines the Windsurf development environment with an Agent Command Center that orchestrates third-party agents — including OpenAI Codex and Claude Agent — via the Agent Client Protocol.

In parallel, Devin Local runs with up to 30 percent more efficiency than previous tools. And because Cognition believes in its own quality, it has offered an AI Productivity Guarantee since June 6: if Devin delivers less engineering value than paid for, Cognition covers the usage costs — up to ten million US dollars (≈ nine million euros), paid out as credits at the end of the annual contract term. Credits, not cash — an important distinction.

This guarantee is an enterprise construct with an annual contract — not meant for SMB end customers. But the signal matters: a vendor that won't guarantee measurable productivity to its customers either lacks confidence in its own product — or has no business case.

Tencent: WorkBuddy and the agent swarm

With WorkBuddy Enterprise Edition and its accompanying Agent Suite, Tencent unveiled a system that lets multiple AI agents work as a team. A OneID system connects Tencent Docs with LearnShare, and MCP integration brings together Jira, GitHub, Google Drive, Gmail, Notion, and Slack.

The special part: tasks can be triggered remotely via Slack, Telegram, Discord, and WeChat. Tencent VP Liu Yi puts into words what many SMB owners know from their daily experience: "N super individuals don't automatically become a super team." That statement doesn't just apply to large corporations. It applies to every trades office where everyone does everything and no one gets anything done.

Vercel: 96% marketing, 93% support — and no more SDR team

Vercel, the platform behind Next.js, shared numbers at SaaStr AI Deploy on June 6 that are worth reading twice. A content agent produces 96 percent of the first marketing drafts — a human finalizes them. A support agent handles 93 percent of all customer inquiries with no human intervention. And a DealOne agent has replaced the internal SDR team: lead qualification now runs fully automated.

CPO Tom Occhino names the philosophy behind it: "We build for ourselves first." Four words that also describe our principle.

Why we know this firsthand

What does this mean concretely for the mid-market?

80 percent code. 96 percent marketing. 93 percent support. These numbers sound like they belong on a conference stage. But the lever translates — for companies that don't have their own AI department and still want to move faster.

Ben is a phone agent for trades businesses and financial advisors. He answers calls, qualifies inquiries, and books appointments — around the clock, even when the office closes at 5pm. 89 percent of Germans reach for the phone first when they have a problem. Ben is there when no human picks up.

Sara is a recruiting agent. She matches applicant profiles against job requirements and produces a shortlist the HR office can review directly. No cold mass of applicants — just the candidates who fit the role.

Tom is a dunning agent. He tracks open invoices, sends payment reminders at the right time, and escalates when necessary. No emotions, no delay. With B2B late-payment interest above 10 percent in the first half of 2026, every open day on an invoice counts.

Pilot: €990 setup, €149 per month. Your agent is up and running in 30 days. If it doesn't deliver, you walk away. No twelve-month contract. No need for a multimillion-dollar guarantee — just one a ten-person business can afford.

How does autonomous AI stay safe and compliant?

Autonomous AI agents aren't a black box. At amazing agents, we document the prompt chain, model version, data flows, and retention periods for every engagement — as required by Articles 5 and 30 of the GDPR. No customer data flows into model training. ManageEngine CEO Rajesh Ganesan puts it clearly for the enterprise market: "Customer data is not used for model training." In writing, in the data protection agreement. The same applies for us — just for the mid-market.

Human-in-the-loop remains mandatory for decisions with legal consequences — Article 22 GDPR and Article 14 of the EU AI Act. No agent of ours makes a decision that a human hasn't approved beforehand. Autonomy means working faster. Not working without control.