Anthropic decouples the brain from the hands —
and the stack follows.

Cloudflare and Anthropic landed the architectural deal of the week. Claude Managed Agents now have a first-class production runtime: agents pick between full Linux microVMs (Cloudflare Containers) for long-running, language-flexible work, or V8 isolates with millisecond cold starts for high-fanout tool calls. A Workers control plane spins a fresh sandbox per agent session, egress runs through Cloudflare Mesh with post-quantum-ready encryption and customer-defined policy, and audit trails plus session recordings ship by default. Anthropic's own framing — "decouple the brain from the hands" — is the strategic line: the model provider runs the loop, the network provider runs the execution, and the customer keeps the data. It's the cleanest answer yet to "where does an autonomous agent actually run code," and the first time the runtime layer has been treated as a separate product from the model.

Five rapid releases in three days, past the prior 141–145 cycle. The headlines: the old /simplify got renamed to /code-review with optional effort levels (low/medium/high → max), GFM checkbox rendering landed in v2.1.149, and an allowAllClaudeAiMcps managed setting now lets enterprises auto-load claude.ai cloud MCP connectors. The practically useful change is the new /usage breakdown — Claude Code now itemises spend across skills, subagents, plugins and MCP-server calls inside a session, the first time a coding- agent CLI surfaces where the tokens actually go. It lands the same week MCPSpend ships the same idea as MCP middleware; together they make "observability for the agent harness" a real category, not a promise. Pinned background sessions also now stay alive when idle and restart in place for updates.

An Apache-2.0 Python proxy from ajsai47 (trending past 200 stars in a week) that sits between Claude Code and any OpenAI-compatible provider — DeepSeek, Groq, Ollama, OpenRouter, NVIDIA NIM, LM Studio, your own local model — and keeps the UX intact. The harness/model decoupling thesis is now a visible product category (backdoor, cc-switch, oh-my-openagent), and the strategic question for Anthropic shifts from "can we keep developers in the API" to "is the harness or the model the thing we're selling." The fact that customers increasingly want the Claude Code UX even when the model isn't Claude is the most important signal here. Both ship for now; the moment one outgrows the other, the answer becomes obvious.

Cole Medin (coleam00) shipped a B2B-helpdesk monorepo on Monday that demonstrates the full Anthropic-blessed Claude Code stack on one real codebase: hierarchical CLAUDE.md files (root + per-service), SessionStart and self-improving Stop hooks, glob-scoped domain skills, a Pyright LSP wired into the agent, a codebase-search MCP server, a read-only "explorer" subagent, and the whole thing packaged as a distributable helpline-ai-layer plugin. Five services and two shared packages, all wired together. It's the cleanest assembled example we've seen of what "Claude Code in a large codebase" actually looks like at the architecture level, and it arrives as Anthropic is publishing its own best-practice guides. Reference architectures for the harness layer are now a thing teams clone.

Accel led a $75M Series A for Viktor — an AI coworker that lives inside Slack and Microsoft Teams — with Slack cofounders Stewart Butterfield and Cal Henderson, Synthesia CEO Victor Riparbelli, and execs from Google DeepMind, Figma and ElevenLabs participating as angels. Warsaw- and Munich-based, founded in 2023, public for ten weeks; $15M annualised run rate and 2,000+ organisations signed. It's the largest Series A ever raised by a Polish-founded company, but the dollar value is the point: the agent surface — the place a user actually talks to the agent — is now its own venture-fundable category, separate from the runtime (Cloudflare), separate from the model (Anthropic), and separate from the data layer (Tribal, below). The stack is splitting, and capital is sorting it out one slice at a time.

Team8 led a $10M seed for Tribal — founded by Yoav Kolodner (ex-Salesforce VP of Engineering), Yakir Daniel (NetApp / Huawei) and Lior Sidi (ex-Wix AI) — to ship a "Metadata Fabric" that maps every object, automation, permission, dependency and business rule inside Salesforce, ServiceNow, SAP, NetSuite and Workday into something an agent can actually reason about. DYDX Capital and Salesforce-ecosystem founders joined as angels. The thesis is sharp: enterprise agents fail in production not because the model is weak but because the metadata layer underneath is missing, and stitching it together per-customer is what eats the proof-of-concept. If Sierra and Decagon are the chat surface on top of enterprise SaaS, Tribal is betting that the layer underneath them is the durable position.

A small TypeScript proxy (EU-hosted on Hostinger) that wraps any MCP server and attributes per-tool spend across Cursor, Claude Desktop, Windsurf and VS Code — by tool, project, and customer. It sits in the same lane Claude Code's v2.1.149 /usage breakdown carved out at the harness layer, but pointed at the MCP-server side of the wire instead. Two independent products converging on "who is racking up MCP cost" in the same week is a real signal: the moment MCP became plumbing was also the moment enterprises started asking who pays the bill. Free tier 25k calls/month; the EU hosting matters more than it looks for anyone running agents against customer data in regulated industries.

A Python MCP server from norika1207 exposing mechanistic-interpretability primitives — anchor dimensions, per-layer functional fingerprints, cross-architecture equivalence — across 23 models and 13 architecture families as tools any agent can call (mercury_anchor_dims, mercury_layer_fingerprint, mercury_cross_arch_equivalent). The point isn't that current coding agents need it; it's that interpretability research, traditionally a paper-and- notebook field, just got an MCP interface, which means agents can start reasoning about why their own model behaves the way it does without a human in the loop. Small repo, conceptually new rung — the next wave of evaluation skills will consume mercury-mcp under the hood.

Marco Porcellato shipped an Apache-2.0 Python daemon that operationalises Andrej Karpathy's "LLM-Wiki" idea: a local-first agent that semantically indexes your Logseq vault, heals broken wiki-links, consolidates entities, and auto-splits dense blocks — directly editing the Markdown files, no cloud, no Logseq HTTP API. Three configurable risk tiers (Safe / Augmented / Surgeon), optimistic concurrency control to survive concurrent human edits, optional Claude Desktop integration over MCP, and a local "Sovereign UI" dashboard on port 8500. PKM tools just got their first credible autonomous-agent integration — and arguably the right pattern for it (local-first, opt-in escalation, plain Markdown).

A Rust daemon from tuchg (~158 stars) that delivers notifications and approve-from- phone for local Codex, Claude Code or Pi sessions over Telegram and WeChat — with the explicit tagline "no hooks, no skills, no MCP, no project changes." Scan a QR code, the agent gets a remote control. Same problem space as Claude Code Routines and Codex remote-control, but from the local side and provider-neutral — the agent stays where it was, only the human is freed to walk away from the laptop. The pattern is converging from both ends now: cloud platforms ship server-side lifecycle, the community ships client-side bridges, and the harness in the middle stays untouched.

aref-vc shipped a Claude Code skill (~174 stars in three days) that auto-loads from ~/.claude/skills/ and converts any chart request into a Tufte-compliant visualisation, with HTML/SVG and React output presets and a defended quote from Envisioning Information: "the principles of information design are universal — like mathematics — and are not tied to unique features of a particular language or culture." Tiny repo, sharp idea: design taste, distilled to a skill, is portable across every project and every harness. Expect the same pattern for typography (Bringhurst), interface design (Tognazzini), and prose (Strunk-White) in the next month — taste itself is becoming a packaging format.

charliehzm shipped an open-source "harness engineering" framework for medical AI coding: a 12-step SOP plus 5-step micro-workflow, 23 skills (5 compliance, 16 general, 2 aliases), 8 medical-grade MCP servers, Presidio PHI redaction wired in with a Chinese-medical recognizer and 31 contextual rules, and a hard policy that L4 PHI never enters a prompt directly. Apache-2.0 code, CC BY-SA 4.0 docs. It's the cleanest open-source analogue we've seen to Anthropic's Claude-for-Legal — same pattern (regulated vertical, dual-jurisdiction compliance), but built as a harness configuration rather than a vendor product. The "agentic for X" question is increasingly going to be answered by harness configs, not new SaaS.

MobiusQuant shipped an Apache-2.0 skill (~144 stars) carrying 964 trading-concept cards distilled from 130 ICT/SMC teaching videos — identification rules, trading implications, common mistakes, related concepts — plus real-time market data and 60+ indicators. The interesting part is the install matrix: a single --platform flag drops it into Claude Code, Codex, OpenClaw or Hermes with self-contained packaging. Cross-harness portability is no longer a value-add; three serious vertical skill packs landed this week (Tufte for charts, medharness for healthcare, OpenMobius for trading) and all of them ship as multi-platform from day one. The implicit assumption: the harness is the OS, the skill is the app.