Skip to content

Code with Claude 2026 — feature mapping for Etz Chaim AI

This document maps every Anthropic 2026 feature announced at or before Code with Claude SF (May 6, 2026) to its role in Etz Chaim AI. Use this as the reference when deciding whether to wire a new Anthropic feature in.

Last sync: May 11, 2026 (3 days after Code with Claude SF keynote)

The Code with Claude SF 2026 announcements

Held at SVN West, San Francisco, May 6, 2026. Three thousand-ish attendees in person. Tracks: Research · Claude Platform · Claude Code.

Headline announcements (with Etz Chaim mapping)

Announcement Etz Chaim impact Sprint
Claude Code 5h rate limits doubled for Pro/Max/Team/Enterprise Maintainer can run longer improve-loops 1
Peak-hours throttling removed for Pro/Max Nightly daemon runs reliably 1
Opus API rate limits considerably raised Adversarial probes can run in parallel 1
300+ MW SpaceX/Colossus 1 partnership · 220 000+ GPUs Backend capacity for Routines (infra)
Dreaming → research preview Memory curation for epistememory faculty 8
Outcomes → public beta Rubric grader per faculty 8
Multiagent orchestration → public beta 11 malakhim as parallel sub-agents 8
Webhooks for Managed Agents Drift alerts → Telegram 7
Harvey case study: 6× task completion with Dreaming Validation of our memory direction (proof point)
Wisedocs: −50% review time with Outcomes Validation of rubric direction (proof point)
Netflix: hundreds of parallel builds via Multiagent Pattern for our 11 adversaries 8
Q1 2026: 80× annualized revenue growth Strategic context (don't bet against the stack)
API volume: +70× YoY Same
Avg dev uses Claude Code 20 h/week Validates the workflow we're standardizing on
4% of public GitHub commits are Claude Code authored Distribution surface for OSS visibility

Anthropic feature catalog → Etz Chaim mapping

Updated table, sorted by sprint priority.

Sprint 1 — Zero-architecture wins

Feature Etz Chaim use Reference
Prompt caching (explicit breakpoints) Cache the 36 086-line spec corpus → ~90% cost reduction on improve-loop docs
Claude Opus 4.7 (released Apr 16, 2026) Upgrade from 4.6 on deep facultés (causal, dissensus, insight) release notes
Adaptive thinking Replace manual budget_tokens on facultés docs
Batch API Nightly daemon runs as batch (-50% cost on top of caching) docs
Citations API Native E1–E6 label mapping docs
Advisor tool (beta advisor-tool-2026-03-01) Sonnet executor + Opus advisor for the dispatcher in triage subagent Anthropic news

Sprint 2 — Skills (the Barry Zhang / Mahesh Murag thesis)

Feature Etz Chaim use Reference
Skills standard (agentskills.io) 13 SKILL.md files for 10 facultés + 3 utility agentskills.io
Progressive disclosure Metadata up-front, full content on demand (Anthropic's stated pattern) Barry Zhang talk
disable-model-invocation: true frontmatter Use on etz-spec-lookup (manual-only skill) Skills docs
Skills marketplaces Publish to skills.sh, skillsmp.com, agensi.io, buildwithclaude.com, awesome-agent-skills skills.sh
Cross-tool skills Same skills work on Codex CLI, Cursor, Antigravity (35+ platforms) Agent Skills explainer

Sprint 3 — Subagents (Boris pattern)

Feature Etz Chaim use Reference
Subagents (/agents/*.md) 5 subagents: triage, verify-spec, improve-loop, spec-auditor, doctor Claude Code docs
isolation: worktree frontmatter On improve-loop and verify-spec (Boris pattern) Boris's worktree thread
/batch slash command Adapt as /batch-migration for spec migrations Boris's /batch announcement
/simplify slash command Adapt for post-mutation cleanup (built-in Claude Code)
Plan Mode (Shift+Tab×2) Mandatory for all spec mutations Boris thread
Verification loop (verify-spec) The 2-3× quality boost insight VentureBeat manifesto

Sprint 4 — Hooks, Sandbox, Auto Mode, DevContainer

Feature Etz Chaim use Reference
Hooks (26 lifecycle events in v2.1.116) 8 hooks: PreToolUse aggregate-write guard, PostToolUse format, Stop verify-app Hooks docs
Sandbox (Seatbelt/Landlock) OS-level isolation for the daemon Sandbox docs
Auto Mode (Sonnet 4.6 classifier) Replace --dangerously-skip-permissions for the daemon Auto Mode blog
Prompt-injection probe Screen Sefaria-sourced content (part of Auto Mode)
DevContainer One-line install via VS Code / Codespaces DevContainer docs
Trail of Bits hardened devcontainer Reference firewall pattern GitHub
/loop & /schedule Boris's autonomous workflows Boris's tips
--worktree flag Each adversary in its own worktree Worktree announcement
statusline custom Show context %, model, branch, cost, active rectifiers statusline docs
/insights + cleanupPeriodDays: 365 Longitudinal flywheel (Shrivu Shankar pattern)

Sprint 5 — MCP server (etzchaim-mcp)

Feature Etz Chaim use Reference
MCP standard (AAIF) Build etzchaim-mcp (15 tools) MCP spec
Tool Search (auto-defer, 95% context reduction) All etzchaim-mcp tools default-deferred Tool Search docs
alwaysLoad: true annotation On diagnose_faculty (critical, no defer) docs
anthropic/maxResultSizeChars annotation Cap result size to 50 000 chars (MCP server best practice)
Tool description optimization Use the tool-testing agent pattern (40% latency reduction) Anthropic eng
Parameter naming (Ken Aizawa rule) spec_id not id, faculty_name not name Writing tools for agents

Sprint 6 — Plugin distribution

Feature Etz Chaim use Reference
Claude Code Plugins (.claude-plugin/plugin.json) Bundle skills + agents + commands + hooks + MCP config Plugins docs
buildwithclaude.com marketplace Primary distribution channel buildwithclaude
skills.sh (Vercel) Cross-tool skills installation: npx skills add yohanpoul/etz-chaim-ai skills.sh
mcp.so MCP server registry mcp.so
Glama MCP server discovery Glama
agensi.io Security-scanned skills agensi.io
awesome-agent-skills Community list GitHub

Sprint 7 — Cloud asynchrone (Routines + Channels + Webhooks)

Feature Etz Chaim use Reference
Claude Code Routines REPLACES local daemon for nightly improve-loop TheNewStack launch
Channels (research preview) Telegram alerts for drift / mutation approval Channels docs
--channels permission relay Approve tool calls from phone (Claude Code v2.1.x)
Webhooks for Managed Agents (whsec_) At-least-once dedup via Hookdeck Hookdeck guide
Claude Code Action GHA PR-from-anywhere via @claude mention claude-code-action
OpenTelemetry expansion in Cowork Trace nightly routines into Grafana (Code with Claude 2026 announcement)

Sprint 8 — Managed Agents (the long bet)

Feature Etz Chaim use Reference
Managed Agents (anthropic.beta.agents.create) Persistent agent for Etz Chaim Managed Agents docs
Memory tool + stores Replace local epistememory storage Memory docs
Dreaming (research preview) Nightly memory curation, review-human-required Dreaming announcement
Outcomes (public beta) Rubric grader per faculty in separate context Outcomes blog
Multiagent orchestration (public beta, managed-agents-2026-04-01) The 11 malakhim as parallel specialist sub-agents Multiagent docs
Persistent events Trace every step in Claude Console (Managed Agents feature)
Vault credentials background refresh mcp_oauth secret rotation (May 2026 release notes)

Sprint 9 — Evangelism & v1.0

Feature Etz Chaim use Reference
Claude Certified Architect Certification track Claude Certified
Code with Claude London (May 19) / Tokyo (June 10) Talk submissions Code with Claude
Claude for Open Source Program 6 months Max free for OSS maintainers (deadline June 30, 2026) Apply

Features we explicitly DON'T use (and why)

Feature Why we skip
Cowork plugins Cowork is for knowledge workers, not OSS infrastructure
Claude in Excel/PowerPoint/Word/Outlook Not relevant to Etz Chaim's core use case
Finance agent templates Domain-specific to financial services
Computer Use (full) Overkill for our needs; the Chrome extension on the Flask dashboard is sufficient
Claude Security We'll use Semgrep + dependency scanning on etzchaim-mcp, but not the full Claude Security product

Reference index

Boris Cherny (creator of Claude Code) — workflow

Anthropic engineering — foundational papers

Anthropic thesis — Barry Zhang & Mahesh Murag

Code with Claude SF 2026

Standards