Boris Cherny: After Claude Code, Programming Becomes Agent ManagementBoris Cherny：Claude Code 之后，写代码正在变成"管理 Agent"

Boris Cherny argues Claude Code has shifted his own job from writing code to orchestrating hundreds of autonomous agents via cron-triggered "Loops," and that this same shift will spread from individual engineers to entire organizations.

His larger claim is that AI erases two of Hamilton Helmer's seven competitive moats — switching costs and process power — while leaving network effects and scale economies intact, implying a wave of small startups can now out-compete incumbents from scratch.

The founding story matters because it inverts standard startup logic: Anthropic Labs built Claude Code explicitly for a model generation that didn't exist yet, accepting zero traction for six months. This is only replicable by an organization that controls both the model roadmap and the product — a structural advantage independent teams building on the API cannot match, since they must build for models as they exist today, not as they will be in six months.

Cherny's "coding is solved" claim is real but narrowly scoped. He explicitly chose TypeScript and React in 2024 because they were the most heavily represented languages in training data — an engineering decision optimized for model competence, not for the codebase's actual needs. The claim's generalizability hinges entirely on whether frontier models close the gap on out-of-distribution stacks (legacy C++, SAP ABAP, embedded systems) as fast as they closed it on React.

The workflow shift to Loops is the piece's most concrete and verifiable claim: dozens of cron-triggered tasks running unattended, escalating via Slack only when uncertain. This maps directly onto Anthropic's Routines product launch, which moves the same pattern server-side — a signal that Anthropic is productizing its own internal dogfooding, and that pricing for always-on agent scheduling is coming.

The Seven Powers argument is the highest-stakes claim and doubles as Anthropic's own strategic bet: Cowork explicitly wagers that switching costs will collapse and desktop agents can take over enterprise workflows. But enterprise SaaS switching costs live mostly in compliance audits, contract terms, and procurement inertia — not migration friction — so a model that can move your data doesn't necessarily move your vendor relationship. The $285B software-stock selloff in February 2026 shows markets pricing in the risk, not that the thesis has been proven.

The printing-press analogy is rhetorically effective but numerically loose: Cherny cites 15th-century European literacy at ~10% and modern global literacy at ~70%, versus scholarly estimates of ~25-30% and ~90% respectively. The direction of the argument — deprofessionalization accelerating content/software production by orders of magnitude — survives the correction, but the piece flags an omission: the press also triggered censorship regimes and religious wars, a parallel to software's coming wave of AI-generated malware and deepfakes that Cherny doesn't address.

The prediction that safety scaffolding (prompt-injection defenses, permission modes, human-in-the-loop) will matter less as models improve is directly contradicted by a documented incident: in April 2026, a Claude Opus 4.6-driven coding agent deleted a production database and its backups, reported by the Guardian. Anthropic's own Opus 4.7 release notes describe safety as "similar to 4.6, not fully solved" — evidence that the erosion of external control layers Cherny anticipates hasn't yet materialized in practice.