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

Claude Code engineer Daisy Hollman explains how to give agents the information and tools they need without filling their limited input space with irrelevant material. She connects that design problem with supervision: as more agents work at once, people need quick ways to see what each has done and where it needs help. Her examples come from the Claude Code team and her personal setup, rather than an organization-wide adoption study.

Key Points Covered

• Audit where engineering work actually lives: Hollman suggests trying to work for a day without leaving the terminal. Every manual trip to team chat, CI, production dashboards, internal documentation, or design history identifies information or an action the agent cannot yet handle. Broader access may close that gap, but the talk does not specify least-privilege, privacy, or governance controls for those systems 15:09–20:22(opens in a new tab).
• Push verification into the edit loop: A PostToolUse hook can append existing type-checker, linter, or policy output to the tool result immediately after an edit. Hollman's operative claim is that timely feedback saves the agent from rediscovering the mistake at a later compile step; she does not measure the quality or token effect in this recording 22:20–25:16(opens in a new tab).
• Leave input space for the task itself: A model's context window holds its instructions, tool definitions, CLAUDE.md, files, and results as well as the current task. Hollman explains that prefix caching reuses computation for the unchanged beginning of that input. Replacing an early part can force the service to process what follows again, so removing context is not always cheaper. Her ten-times cost example illustrates the tradeoff rather than establish a benchmark 26:22–33:22(opens in a new tab).
• Load tool details when needed: Every tool exposed through the Model Context Protocol (MCP) adds a name, description, and input schema. Tool search can load names first and defer the details until the agent selects a tool; clear names therefore matter. Where a company already has a command-line tool, Hollman often prefers instructions teaching the agent to use it over building another MCP server 33:22–38:07(opens in a new tab).
• Skills, subagents, and hooks spend context differently: A skill is reusable task guidance: its short description is loaded up front, while its full instructions load when selected. A subagent does work in a separate context window and returns a summary, keeping its intermediate investigation out of the parent's conversation. A hook runs ordinary code after an event, such as a file edit, and supplies output only when triggered. These mechanisms reduce different kinds of input overhead; none makes all relevant information free 38:07–49:16(opens in a new tab).
• Separate maintained context from model-curated memory: Both may be text files, but Hollman reserves context engineering for reusable, deliberately maintained instructions and separates it from model-curated memory 48:14–49:16(opens in a new tab).
• Isolate persistent agents and make them easy to recognize: Hollman describes her personal setup of named, colored, long-lived agents parked in separate Git worktrees, each responsible for tracking upstream main and its own scratch state. She says identity and color reduce her context-switching burden, but provides no comparative productivity or quality measurement and does not present the setup as an Anthropic-wide standard 49:16–53:56(opens in a new tab).
• Coordinate delayed and parallel work with explicit channels: Agent teams can use a message primitive with a leader-and-teammates or peer topology, while /loop can wake a session to recheck delayed CI. Hollman says auto mode combines prompt authorization checks with models and classifiers so these workflows can run without repeated approval; the control design is not independently audited here, and she reports a 10–40% model-dependent cost premium 53:56–58:02(opens in a new tab).
• Summarize agent state for the human bottleneck: Fleet View uses lightweight classifiers to show what multiple sessions did and need, while remote control lets Hollman reconnect to persistent agents from another device. The aim is to reduce the time a person spends reconstructing each session's state before deciding what to do next 56:58–1:00:09(opens in a new tab).

Full video: https://www.youtube.com/watch?v=shZgedW15vg(opens in a new tab)