# Rethinking Claude Code Best Practices — Lydia Hallie, Anthropic

Source: [AI Engineer](https://www.youtube.com/watch?v=vlfCjWuv4es)  
Feed7 permalink: https://feed7.dev/p/rethinking-claude-code-best-practices-lydia-hallie-anthropic-15e6kae  
Published: 2026-11-12T03:26:55.000Z  
Trust: Source Linked (source_linked)

## Why Included

Anthropic’s Claude Code talk shifts best practices toward sandboxed execution and enforced verification as agents run longer and with less direct supervision.

## Source Summary

The talk describes longer-running Claude Code sessions, with auto mode handling some approvals and sandboxes limiting shell-command access. In a cited study, auto mode blocked **89%** of dangerous commands, while people caught **about 14%**.

## Practical Implication

If you run coding agents unattended, define what commands can reach, isolate sessions where possible, and make tests or other checks gate completion. The speaker also recommends trimming setup instructions that newer models no longer need and using hooks for checks that must run.

## Agent-Ready Context

The talk describes longer-running Claude Code sessions, with auto mode handling some approvals and sandboxes limiting shell-command access. In a cited study, auto mode blocked **89%** of dangerous commands, while people caught **about 14%**.

If you run coding agents unattended, define what commands can reach, isolate sessions where possible, and make tests or other checks gate completion. The speaker also recommends trimming setup instructions that newer models no longer need and using hooks for checks that must run.

Auto mode still let roughly **one in ten** dangerous commands through in that study. Sandboxing covered shell commands and their child processes, but not Claude’s own files, MCP servers, or hooks; verification quality still depends on the checks you configure.

## Connected Context

Feed7 judgment across 1089 accumulated Signals:

This frames unattended coding-agent safety as layered containment and verification, not as approval automation alone. The reported auto-mode block rate leaves residual risk, and the stated sandbox boundaries exclude files, MCP servers, and hooks, so those paths need separate controls.

- [We let an AI agent execute Bash and lived to talk about it — Sarah Sanders, PostHog](https://feed7.dev/p/we-let-an-ai-agent-execute-bash-and-lived-to-talk-about-it-sarah-sanders-0pgksvk) — PostHog describes layered scanning and deterministic controls, reinforcing the need for safeguards beyond command approval while also relying on sandboxing and bounded authority.
- [apache/maka](https://feed7.dev/p/maka-01vaj44) — Maka packages a sandbox with permissioned execution, durable logs, and recovery, offering runtime controls relevant to the talk’s unattended-session risks.
- [In Code They Act, In Proof We Trust — Erik Meijer, Leibniz Labs](https://feed7.dev/p/in-code-they-act-in-proof-we-trust-erik-meijer-leibniz-labs-1lyno2y) — The proposed proof-gated execution makes tool side effects subject to a separate safety check, complementing the talk’s emphasis on controls outside approval automation.

## Context Map

- Layer: agent
- Domains: coding, security
- Topics: harness-engineering, sandboxing, agent-reliability

## Uncertainty

- Auto mode still let roughly **one in ten** dangerous commands through in that study. Sandboxing covered shell commands and their child processes, but not Claude’s own files, MCP servers, or hooks; verification quality still depends on the checks you configure.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.
