How Anthropic uses Claude Code: Agentic Software Engineering at Scale - Daisy Hollman

How Anthropic uses Claude Code: Agentic Software Engineering at Scale - Daisy Hollman

by Daisy Hollman

Applied Sciences & Engineering Computing & Cybersecurity UMComputer programmingUMZSoftware Engineering
🎙 Daisy Hollman 👥 227K 📅 August 11, 2026 ⏱ 60 min 👁 6K 📄 expert opinion 🧭 2026-08-13
Available in: English (current) Français

Keywords

Claude Codeagentic programmingcontext windowpluginssoftware engineering

Summary

Daisy Hollman, an engineer at Anthropic, presents a talk at NDC Copenhagen on how Anthropic uses Claude Code for agentic software engineering at scale. She begins by distinguishing chatbots from agents, explaining that agents use tool calls to perform actions and iterate. She highlights the primitive nature of current agent tools, such as the edit tool, which is essentially find-and-replace. She discusses the exponential growth in agent capabilities, citing a chart from METR showing that the time horizon for tasks agents can complete with 50% success is doubling every four months, and mentions Mozilla’s April bug fixes as a practical example. The core thesis is that for Claude to do your job, it needs access to the same information and tools you have. She emphasizes the importance of customization through in-context learning, as model weights are frozen. The context window is the key constraint, and its size has stagnated at around 1 million tokens, making context engineering a critical discipline. She introduces plugins as a way to inject knowledge and tools, and discusses post-tool-use hooks for providing immediate feedback. She argues that software engineering is becoming about teaching agents, and that context engineering is the new core skill. The talk concludes with insights into how Anthropic uses Claude Code internally, including agent teams and scaling practices.

216 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the practical challenges of scaling agentic software engineering, particularly around context window limitations and the need for customization. Hollman’s argumentation is coherent and grounded in her direct experience at Anthropic. She effectively uses analogies (e.g., ED vs. VS Code) to illustrate the current state of agent tooling. The claim that context engineering is becoming the core of software engineering is compelling and well-supported by examples. However, some arguments rely on anecdotal evidence and projections (e.g., Moore’s law of agents) that are not rigorously substantiated.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates strong technical rigor, with clear explanations of tool call mechanics and plugin design. Hollman references METR’s chart and Mozilla’s data, but these are not formally cited with URLs. The title accurately reflects the content, focusing on Anthropic’s use of Claude Code. The talk is an expert opinion rather than a peer-reviewed study, so the scientific rigor is moderate. The description provides links to NDC conferences but no direct references to the mentioned data sources.

182 words

Title / Content Match

The title accurately reflects the content: the talk focuses on how Anthropic uses Claude Code for agentic software engineering at scale, covering context engineering and plugin design.

Quality & Reliability

8/10

The speaker is a senior engineer at Anthropic with deep technical expertise, providing an insider perspective on Claude Code's design and usage. The talk is grounded in practical experience and references real-world data (e.g., Mozilla's bug fixes). However, it is largely anecdotal and lacks peer-reviewed sources, and some claims (e.g., Moore's law of agents) are presented without rigorous evidence.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • NDC Conferences — Conference organizer for the talk.
  • NDC Copenhagen — Specific conference where the talk was recorded.

Concurring Sources

Contribution & Novelties

The talk provides an insider perspective on how Anthropic designs and uses Claude Code for large-scale agentic software engineering. It introduces the concept of context engineering as a core discipline, and discusses plugin design and post-tool-use hooks as mechanisms for customization. The emphasis on the stagnation of context window sizes and the need to optimize within that constraint is a novel framing.

Pour aller plus loin :

  • In-context learning — Relevant to the discussion of customization via text.
  • Tool use in LLMs — Background on tool calling.
  • Context window — Explanation of the concept.

94 words

Radar Profile

The radar profile shows high scores in information quantity and quality, reflecting the speaker's expertise and the depth of content. The technical level is high, but the reliability is slightly lower due to the anecdotal nature of some claims. This suggests a talk that is informative and technically rich but may require additional verification for some assertions.

Reliability 7/10