← Back to Claude Code setup

Claude Code Landscape — March 2026

Survey of the Claude Code ecosystem: power-user patterns, multi-agent orchestrators, memory tools, autonomous execution methods, and key people.

Claude Code Landscape — March 2026

Compiled from X/Twitter via Grok API + Exa web search. Last updated 2026-03-25.

How Power Users Are Using Claude Code

The 4-Layer System

CLAUDE.md → Skills → Hooks → Agents. This is the consensus architecture for serious Claude Code usage. Each layer adds structure without removing flexibility.

Day-to-Day Patterns

  • CLI for code review, PRs, deploys, refactoring via slash commands
  • Config via .claude/ (CLAUDE.md for architecture/rules, skills/agents for auto-triggers)
  • /clear between tasks to reset context
  • Memory files for compounding rules across sessions
  • Reports saving 3-4 hours per project but requiring domain knowledge

What People Love

  • "Easily the best price-to-usage AI coding setup"
  • Auto-spawns agents (reviewer/security), self-improves
  • 1M token context fixes previous limits
  • From experiment to "tool I cannot work without"

What Frustrates Them

  • "Shocking mistakes" — overly complex code, ignores style, breaks prod
  • Needs 100% supervision; "confidently wrong"
  • Slow TUI (10min no feedback), permission popups
  • Mediocre without .claude/ config habit
  • "Not sure delivering more value" vs manual for some users

Multi-Agent Orchestrators

Agent Teams (Built-in, Experimental)

  • One lead coordinates, teammates communicate directly
  • Shared task list with self-coordination
  • 3-5 teammates, 5-6 tasks each recommended
  • --dangerously-skip-permissions flows to all teammates
  • Enable: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Gas Town (Steve Yegge)

  • Dominant third-party orchestrator, OSS
  • Manages swarms of Claude Code agents
  • Called "the new IDE for 2026"
  • Powerful but overwhelming ("AI brain fry")
  • Steve Yegge runs 3 concurrent Claude Max accounts to feed it
  • git clone https://github.com/steveyegge/gastown

Multiclaude

  • Simpler than Gas Town
  • Better for "give long prompt and walk away"
  • Better for team usage/review

OpenClaw (formerly Clawdbot)

  • Always-on AI daemon, runs on your machine or VPS
  • Full system access: files, email, calendar, web, apps, shell, sub-agents
  • Connect via WhatsApp/Telegram/Discord/Signal
  • Uses API keys (not your Claude subscription)
  • Open source, curl -fsSL https://get.openclaw.ai | bash
  • Created by Peter Steinberger (@steipete), now community-driven
  • Renamed from Clawdbot after Anthropic trademark C&D
  • Use cases: inbox management, trading bots ($50→$828 Polymarket), 550 UGC videos/day

Other Orchestrators

  • Beads — agent orchestrator, used alongside Gas Town
  • KERNEL (Aria Han) — plugin that auto-classifies task complexity, routes to agent tiers, stores research/plans in shared DB
  • GStack / Overstory / Mission Control / DeerFlow — various harnesses

Memory & Persistent Context

claude-subconscious (Letta)

  • Background agent that watches all sessions, builds persistent memory
  • Auto-injects relevant memories before each prompt
  • First prompt: full memory block. Subsequent: diffs only (near-zero token cost)
  • Runs on Letta's cloud (api.letta.com), default model is free (zai/glm-5)
  • Install: /plugin marketplace add https://github.com/letta-ai/claude-subconscious then /plugin install claude-subconscious
  • Needs LETTA_API_KEY from app.letta.com

OpenMemory (Mem0)

  • Portable memory across Claude Code/Desktop/Cursor/Copilot
  • Hooks into session start/end/compaction
  • Extracts decisions and anti-patterns into long-term memory
  • Web dashboard at app.openmemory.ai
  • Install: /plugin add mem0ai/claude-code-plugin

Built-in Memory System

  • MEMORY.md files in .claude/projects/ directories
  • Project-scoped, loaded automatically per working directory
  • Manual but reliable

Autonomous Execution Methods

Auto Mode (Released March 2026)

  • Shift+Tab or claude --enable-auto-mode
  • Safeguards classifier auto-approves low-risk actions, prompts for high-risk
  • Middle ground between manual approval and full skip

--dangerously-skip-permissions

  • Skips ALL prompts, no safety net
  • Required for true unattended operation (ralph loop, overnight builds)
  • Risk: prompt injection from malicious repo files (60k+ public repos vulnerable)

Ralph Loop

  • Plugin that restarts Claude Code when it exits
  • /ralph-loop to start
  • Combined with skip-permissions = true autonomy
  • Reports of 27 hours / 84 tasks in one run

Auto-Compact Configuration

  • CLAUDE_CODE_AUTO_COMPACT_WINDOW — set context capacity for compaction (e.g., 200000 on 1M model)
  • CLAUDE_AUTOCOMPACT_PCT_OVERRIDE — percentage trigger (e.g., 80 = compact at 80% of window)
  • Prevents quality degradation on long sessions

Remote Control

  • claude rc — control your local session from phone/browser
  • Session runs on your machine, phone is just a window
  • Your CLAUDE.md, MCPs, local files all stay available
  • Laptop must stay on (unlike cloud)
  • Max plan (Pro coming)

Claude Code on the Web (Cloud)

  • Runs on Anthropic's sandboxed Linux VM at claude.ai/code
  • No access to local files/MCPs/config
  • Laptop can be off
  • Pro or Max plan

Key People to Follow

  • @steipete — Peter Steinberger, OpenClaw creator
  • @steveyegge — Steve Yegge, Gas Town creator
  • @bcherny — Boris Cherny, Claude Code creator at Anthropic
  • @noahzweben — Claude PM
  • @MatthewBerman — AI tool reviews
  • @PrakashS720 — Claude Code power user tips
  • @trq212 — Claude Code workflows
  • @ariaxhan — Aria Han, KERNEL plugin creator

Viral Takes (March 2026)

  1. Desktop control preview (123k likes) — Claude uses your apps/browser/spreadsheets
  2. Channels release (25k likes) — message Claude from phone via Telegram/Discord
  3. Pentagon hack (7k likes) — flagged $4.2B overpays, drafts bids
  4. Matt Shumer "Something Big Is Happening" (86.1M views) — saved in Workspace

Anthropic Harness Design (March 24, 2026)

Source: https://www.anthropic.com/engineering/harness-design

Three agents: Planner → Generator → Evaluator

Key findings:

  • Self-evaluation doesn't work — separate the builder from the judge
  • Opus 4.6 runs coherently for 2+ hours with auto-compaction
  • Sprint decomposition was scaffolding for weaker models — unnecessary on Opus 4.6
  • Communication via files, not messages
  • "Find the simplest solution possible, only increase complexity when needed"

Full details in CLAUDE.md and Autonomous Build Workflow.md.