Claude Code Digest — Aug 10–Aug 13

The biggest shift this week: Claude Code is moving from a prompt box to an execution platform, with policy, sandboxing, and auditability becoming more important than raw model cleverness.

77.4% of skill installs are held by the top 1%

Trending Now

🔥 200-Subagent Ceiling Removed: Run Truly Parallel Workflows
Claude Code 2.1.224 removes the old 200-subagent cap, which changes orchestration from "careful batching" to "scale until your infra hurts." If you’ve been serializing work to stay under the limit, rework your planner to fan out aggressively and use cross-session SendMessage/ListAgents for multi-machine coordination. 🔥 MCP Security Is Now About Cache Keys, Not Just Tool Permissions
The sharpest MCP failure mode this week is cross-scope leakage through bad cache keys. If tenant, role, project, or auth state are not part of the cache key, you can serve the wrong data to the wrong user. Add negative cross-tenant tests now; assume every shared cache is hostile until proven otherwise. 📈 Skills Are a Winner-Take-All Market: Bundle or Die
With Gini at 0.96, skill adoption is extremely concentrated: the top 1% hold 77.4% of installs. That means generic skills are getting buried. Ship niche, workflow-specific skills, bundle them into a coherent package, or target the fast-growing interrogation category where users are actively asking Claude to inspect and explain.

Best Practices

Move hard rules out of CLAUDE.md and into `hooks.PreToolUse`
Before: every session burns context re-reading policy text and still misses enforcement. After: the rule is enforced at tool-time, with less prompt bloat and fewer accidental violations. Guard MCP writes with a GET-before-write diff
Before: empty-payload checks still let no-op writes through, creating false audit entries and noisy automation. After: compare current vs proposed values and only write when the value actually changes. Add auth attributes to every MCP cache key
Before: a shared cache can leak tenant-scoped or role-scoped results across users. After: cache entries are isolated by tenant, role, project, and auth context, closing a real authorization hole.

Tools & MCP

Graft — A semantic code map that persists across sessions via hooks — cuts exploration tool calls and pushed SWE-bench Verified to 66%. @agent-infra/mcp-server-lead-enrichment — Live B2B enrichment for firmographics, technographics, and intent signals — only pays when `confidenceScore > 0.6`, so low-quality matches cost nothing. Octopus Deploy MCP Server — Lets Claude Code inspect environments, query releases, and trigger Kubernetes deployments from `claude mcp add` — reduces onboarding friction and context switching.

Multi-Agent Patterns

Cross-Session Agent Mesh with SendMessage/ListAgents
Use Claude Code 2.1.224’s cross-session messaging to coordinate agents across machines instead of keeping everything in one terminal. This is the first practical path to persistent, distributed agent workflows without rebuilding your own control plane. Subagent Delegation for Exploration
Push open-ended repo exploration into subagents instead of letting the main session wander. That preserves context for the primary task and reduces the classic "I spent 20 turns just looking around" failure mode. Agent Plugins 1.0 Bridge to Claude Code
Convert plugin packages by moving `plugin.json`, renaming `mcp.json`, and swapping two tokens; skills remain usable. This gives teams a migration path instead of waiting for native plugin support.

Community Requests

• Native MCP benchmarking harness with cross-tenant negative tests and cache-isolation checks
• First-class policy/audit primitives for tool-time enforcement instead of relying on CLAUDE.md conventions
• A reproducible evaluation suite for agent plugins, skills, and multi-agent orchestration