Code w/ Claude 2026 — Feature Detail #
- Code w/ Claude SF 2026: Building on the AI exponential (Anthropic, 2026-05-06) — Full official post-event summary. Key additions to last gather’s coverage: Agent View (research preview — single list of all running/blocked/done Claude Code sessions;
claude agents to launch); –plugin-url flag (fetch a plugin .zip from a URL for the current session, enabling ephemeral plugin installs without config edits). The Colossus/SpaceX data center partnership signals Anthropic investing in sustained-compute infrastructure for long-running agents. - Code with Claude 2026: 5 New Agent Features Anthropic Just Shipped (MindStudio) — Practical breakdown of all five: Agent View, Dreaming, Outcomes, Multiagent orchestration, doubled rate limits. Useful detail on Dreaming: the background agent reviews past sessions on a schedule, not on-demand — it’s always running in the background after sessions end, surfacing patterns and improving the memory store without user prompting.
- I Tested 7 Claude Code New Features You Likely Missed (Medium) — Practitioner test covering features released in the 10 days before the event:
claude agents agent view, --plugin-url flag, session-scoped memory, and async hook improvements. Note: Medium source but concrete hands-on detail not available in official docs yet.
Hooks — Deeper Maturity #
- Claude Code: Auto-Approve Tools While Keeping a Safety Net with Hooks (dev.to) — Practical pattern: combine
permissions.allow allowlist for safe tools + PreToolUse hook for conditional approval of borderline operations. Hook receives JSON via stdin; exit 0 = allow, exit 2 = block. Handles cases like “approve git commands but never git push to main.” More surgical than static allowlists or full auto mode. - making Claude Code more secure and autonomous (Anthropic Engineering) — The engineering rationale for auto mode sandboxing. Two-stage classifier: fast initial filter for clearly safe/unsafe tool calls, deeper analysis only for ambiguous cases. The design goal is to reduce human approval interruptions while maintaining a security posture comparable to careful manual review. Introduces .claudeignore as the primary mechanism for defining the agent’s trust boundary.
- GitHub: hesreallyhim/awesome-claude-code (GitHub) — Curated list covering skills, hooks, slash-commands, agent orchestrators, applications, and plugins. A more actively maintained alternative to the existing yurukusa hooks hub; broader scope covers the full Claude Code ecosystem rather than just permission hooks.
Context Engineering & Subagents #
- Effective context engineering for AI agents (Anthropic Engineering) — Anthropic’s framework for context design in production agents. Core principle: agents should maintain lightweight references (file paths, URLs, stored queries) and load data at runtime rather than pre-loading everything into the context window. Contrasts with naive RAG approaches that dump all retrieved content regardless of relevance. Cross-tool: AGENTS.md is cited as the primary mechanism for static context injection at session start.
- Claude Code vs Claude Agent SDK: Which Is for What (Augment Code) — Clear technical distinction: Claude Code = interactive developer tool with its own permission model and UI; Claude Agent SDK = the same agent harness as Claude Code, exposed as a Python/TypeScript library for embedding in applications. The SDK was renamed from “Claude Code SDK” in March 2026. Use Code for interactive work; use SDK for programmatic agent orchestration within larger systems.
- Create custom subagents — Claude Code Docs (Anthropic Docs) — Official documentation for the Agent tool’s subagent model. Each subagent runs in its own context window with a custom system prompt, specific tool access, and independent permissions. The operator pattern: orchestrator receives high-level goal → breaks into subtasks → delegates to subagents → synthesises results. Good reference for understanding isolation model vs. shared-context patterns.
Cross-links #
- Vibe Coding Approaches Anthropic’s context engineering post is the technical foundation for what the AGENTS.md cross-tool standard implements — both are solving the same problem (what does the agent know at runtime) through different mechanisms (runtime loading vs. static injection).
- Vibe Coding Approaches The Claude Agent SDK rename (from “Claude Code SDK”) is a signal that Anthropic is positioning the harness as infrastructure for any agentic application, not just coding tools.
- Emerging pattern: The permission/autonomy dial is now a first-class engineering concern. Auto mode, hooks, allowlists, and the Agent Governance Toolkit (Microsoft) are all solving the same problem from different angles. The design space is clarifying: allowlists for static known-safe patterns, hooks for conditional logic, auto mode for ambient risk-classification. Worth tracking whether these converge into a standard interface.
- Keyword suggestion:
"claude code" "agent view" sessions — Agent View is very new and the practitioner experience docs will appear in the next few weeks.