One Year After “The Emperor Has No Clothes”: Thorsten Ball on Agentic Programming

"The Emperor Has No Clothes": How a Blog Ignited Agentic Programming

In early 2025 Thorsten Ball published the blog post How to Build an Agent: The Emperor Has No Clothes , showing that giving a large model only three primitive tools—read‑file, list‑directory, and ripgrep‑style search—lets it solve code‑base questions without any vector indexes or “Apply” models. The post, released just before Claude Code, became a catalyst for the Agentic programming narrative.

From Batch Changes to Amp: Killing YAML with Natural Language

Ball recounts his experience building Sourcegraph’s Batch Changes, which required engineers to author complex YAML pipelines for multi‑repo modifications. The variability of Go projects (vendor directories, differing go.mod locations) forced a tedious edit‑run‑edit loop, a “bad user experience”. When revisiting the problem, he argued that a large model could replace the entire YAML workflow, eliminating the need for hand‑written configuration.

The insight proved true in Amp: with sufficient token budget, the model runs a continuous loop —calling tools, reading results, deciding next steps—producing solutions that previously needed heavyweight infrastructure.

Amp’s Product Philosophy: Aggressive Subtraction

When asked how Amp differs from other coding agents, Ball says the product serves the “frontier” and deliberately removes features that most teams consider standard. Tab completion, VS Code plugins, custom slash commands, and sub‑agent templates have all been cut because, as models grow stronger, such scaffolding becomes “noise” that merely inflates token usage without improving output quality.

He also explains Amp’s business model: individual users pay only for API token costs, while enterprise customers provide the bulk of revenue. The team refuses contracts to integrate self‑hosted open‑source models, fearing such deals would lock them into a specific stack and hinder rapid model iteration.

Agents and Orbs: The Next Migration in Development Workflows

Amp recently launched “Agents and Orbs”, a system that runs agents on ephemeral cloud machines. When a user issues a command, the platform spins up a fresh VM, clones the repository, installs plugins, and lets the agent operate independently. Users can monitor progress via phone, web, or CLI.

Ball likens this evolution to the shift from single build servers to highly parallel CI runners, predicting that future coding agents will run asynchronously, trigger on events, and scale without a developer watching a terminal.

Is Engineer Experience Still Valuable?

During the interview, Morten noted that heavy reliance on agents can erode a programmer’s grasp of their own code. Ball agrees, distinguishing between low‑impact code (UI scripts, CSS) where AI‑generated solutions are acceptable, and core business logic where deep understanding of data storage, indexing, and traffic patterns remains essential.

He observes that many judgments once reserved for senior engineers—such as the two‑step database migration pattern (add column, double‑write, then drop old column)—are now automatically suggested by models. The irreplaceable skill is the trade‑off judgment built from years of production experience.

Industry Metaphors and the Future of Programming Joy

Ball uses the carpenter and fashion‑designer analogies to argue that while fundamental skills must still be taught, the day‑to‑day manual work can be abstracted away. He stresses that domain knowledge—understanding a company’s core business—will become the scarce commodity, not low‑level coding tricks.

Open‑Source vs. Closed‑Source: An Unstoppable Arms Race

Ball asserts that closed‑source model vendors are in a precarious position. The industry assumes larger models are always possible and that AGI will eventually arrive, prompting a relentless cycle of releasing ever‑stronger models. Because users can switch models instantly, profits from a single model must be reinvested into the next generation, or the vendor risks being overtaken by open‑source alternatives.

He also criticizes government‑mandated restrictions on model access, advocating for “freedom of intelligence” where private companies decide openness without geopolitical constraints.

Compute Scarcity and the Misunderstood AI Bubble

Contrary to the narrative that AI companies are financially fragile, Ball highlights a severe shortage of compute resources. He cites a rumor that a leading model provider considered pausing new user registrations not for cash flow but because they lacked enough GPU capacity.

This mirrors early AWS, where profits were deliberately sacrificed to fund rapid infrastructure expansion. Today, the bottleneck spans raw copper, data‑center construction, and energy supply, even driving up consumer device prices.

Conclusion: Stay Curious and Rediscover the Joy of Coding

Ball ends with a reflective note: maintaining an open mind may reveal new forms of beauty in programming, even as the tools and paradigms evolve dramatically.