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

Claude Code engineer Daisy Hollman presents large-scale agentic software engineering as a context-routing and supervision problem rather than merely code generation. The title should not be read as an organization-wide adoption study: Hollman mostly explains Claude Code primitives, practices from its team, and her personal multi-agent setup, without a rollout sample, comparative workflow, defect results, or business outcomes.

This is an Anthropic first-party conference talk. Its model-capability, cache-cost, product-safety, usage-cost, Mozilla bug-fix, and developer-throughput statements are presenter or vendor reports rather than independent evaluations; several roadmap details are explicitly tentative.

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 (00:15:09–00:20:22(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 (00:22:20–00:25:16(opens in a new tab)).
• Treat context and prefix caching as architecture constraints: Instructions, tool definitions, CLAUDE.md, files, and results all compete with task work. Hollman also argues that dynamically swapping early context can invalidate reusable prefix computation; her ten-times cost illustration and comments about earlier Cursor behavior are explanatory first-party claims, not a published benchmark here (00:26:22–00:33:22(opens in a new tab)).
• Do not expose every enterprise integration as eager MCP context: Every MCP tool contributes a name, description, and schema. Tool search can load names first and defer the rest, but discoverability then depends on informative names; where a company already has a CLI, Hollman often prefers a skill that teaches the agent to use it rather than another server surface (00:33:22–00:38:07(opens in a new tab)).
• Choose a primitive by when its context cost is paid: Skill bodies load on demand but their descriptions remain eager; subagents move work into another context window but still return summaries and require discoverable descriptions; hooks run outside the model context and inject text only after a matching event. Hollman says this is why plugins do not get an unconditional CLAUDE.md primitive, while a session-start hook can still make that deliberate cost explicit (00:38:07–00:49:16(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 (00:48:14–00:49:16(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 (00:49:16–00:53:56(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 (00:53:56–00:58:02(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. Her uncertain recollection that one colleague reached about 1,000 pull requests in a week is an unverified throughput anecdote with no merge, review, defect, or business-outcome definition (00:56:58–01:00:09(opens in a new tab)).

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