Anthropic blinks on Mythos —
and the stack underneath learns to charge.

On May 22, Anthropic published its first formal update on Project Glasswing — and buried the structural news in the same post. Mythos Preview, deployed to roughly 50 trusted security partners, surfaced more than 10,000 high- or critical-severity vulnerabilities in a month, with most partners reporting a 10×-plus uplift in bug-discovery rate over prior Opus-4.6 runs. Cloudflare logged ~2,000 internal bugs (400 classified high/critical) and "fewer false positives than conventional human-led testing"; Mozilla shipped 271 fixes into Firefox 150 traced directly to Mythos. Across 1,000+ open-source projects Mythos flagged 23,019 issues, with the sampled 1,752-finding cohort confirmed at 90.6% and 62.4% rated high or critical. Then the line worth quoting: Anthropic said it looks forward to making Mythos-class models available through a general release "in the near future," conditioned on safeguards that don't yet exist — a clean reversal of April's "no public release" stance. The next day a Mythos 1 toggle briefly appeared in the public Claude Code client (model id claude-mythos-1-preview) before being pulled, and "Access to the Claude Mythos model in Claude Code and Claude Security" surfaced as a string in the client. The most-restricted frontier model in the lab is now publicly on track to ride into the coding agent, and the security team's dashboard, first.

The new enterprise-services joint venture backed by Anthropic, Blackstone, and Hellman & Friedman — unveiled May 4 with $1.5B committed alongside Apollo, General Atlantic, GIC, Sequoia and Goldman — closed its first acquisition: San Francisco-based Fractional AI, founded in 2024 by Chris Taylor, Eddie Siegel and Travis May, and one of the better-known end-to-end AI implementation shops. Terms undisclosed. The interesting part is the unwind: Fractional ended an 11-month partnership with OpenAI to take the deal — exactly the kind of services tie that OpenAI's own May 11 $4B "Deployment Company" was built to lock in. The pattern is explicit now: the model providers no longer trust the consulting layer to be provider-neutral, so they buy it. Sierra and Decagon sit at the chat surface; Tribal is staking the metadata layer; this venture is staking the implementation services that decide which model actually gets deployed inside the Fortune 500.

At Sapphire 2026 in Madrid, SAP turned the entire ERP estate into an agent-first runtime: a unified SAP Business AI Platform that folds together BTP, Business Data Cloud and SAP Business AI under a single SAP Knowledge Graph that gives agents a structured map of every entity and process across a customer's SAP landscape. On top of that platform sits an "autonomous suite" — 50+ domain-specific Joule Assistants across finance, supply chain, procurement, HCM and customer experience, each orchestrating a subset of 200+ specialised agents. Anthropic is named as one of the foundation-model partners powering Joule across HR, procurement and supply chain. SAP also launched Joule Studio (no-code, pro-code and AI-first agent builder on SAP-managed infrastructure) and committed €100M to fund partners building on it. ERP is the back office most of the Fortune 500 actually runs on; the bet is that whoever owns the agent runtime that touches the SAP record of truth owns the deepest moat in enterprise AI.

Kore.ai unveiled the Artemis generation of its agent platform — a serious bet that the right primitive for enterprise multi-agent systems is a compiled, declarative language. The flagship pieces: Agent Blueprint Language (ABL), a declarative IR that standardises how agents, systems and workflows are defined, validated and governed; Arch, an AI-agent architect that translates business objectives directly into production-ready ABL and continuously refines it from production traces; and a "Dual-Brain" runtime that pairs agentic reasoning with deterministic flows through shared memory. Artemis launches natively on Microsoft Azure — integrated with Foundry, Agent 365, Entra ID and Microsoft Graph — with AWS and Google Cloud follow-ons later this year. It's the most opinionated answer this week to "what is the right primitive for multi-agent enterprise systems," and it's the second platform after SAP this week to insist that the missing layer is governance and lifecycle, not yet another framework.

Codex 0.134 stabilised after a four-alpha train (May 22–23). Three changes matter. First, search across local conversation history with case-insensitive content matches and result previews — a quiet but big shift, because it makes the CLI itself a queryable record of every prior agent run on the machine. Second, MCP plumbing goes first-class: per-server environment targeting, OAuth for streamable HTTP MCP servers, and read-only MCP tools running concurrently when they advertise readOnlyHint — every well-written MCP server just got faster by default. Third, extensions and hooks now receive richer context including conversation history and subagent identity, and connector tool schemas preserve local $ref/$defs structures while compacting oversized schemas before exposure. Past the headline features, the schema and concurrency work is the bigger story: the harness is treating MCP like a real wire protocol, not a polite suggestion.

Pydantic AI shipped three v2.0 beta releases (b1→b3) and three v1.x point releases (v1.100→v1.102) in three days — the harness-first rewrite is closing out. The substantive additions in b2/b3: a pending message queue, MCP background tasks (multi-step long-running tool calls that don't block the main run), a new ctx.enqueue API for non-blocking work, model-agnostic XSearch, and Bedrock native-JSON outputs on the v1 line. The whole train was bookended by IPv6 transition-form security normalisation, which lands the same week as a couple of CVE notices in the MCP ecosystem — sober timing. The conceptual move is the same one Codex 0.134 makes: stop treating MCP as a synchronous request/response shim and start treating it as a real, durable, multi-tool runtime with background primitives. The two reference open-source agent frameworks are converging on the same plumbing in the same week.

Doorman11991 shipped a terminal-native coding agent explicitly architected around the constraints of consumer-grade local models — sweet spot 8B–35B parameters, with the polyglot-mini suite as the in-tree benchmark. The design choices are the story: context budgeting that aggressively trims tool output before it hits the model, patch-first editing (the agent emits surgical diffs, not full-file rewrites), and TODO-driven planning so multi-step work doesn't depend on the model holding the whole plan in head. The repo's own README is honest that models below ~4B "struggle with multi-step tool use" and bigger than ~35B "don't benefit" from the optimisations — which is the most useful framing of "where local models actually live" this side of Forge. Eighteen releases in nine days, climbing fast. The harness/model decoupling thesis now has its small-model wing.

The open-gsd team picked up "Getting Shit Done" after a community trust split and shipped a 1.1k-star rewrite in five days. The thesis is sharp: long agent sessions degrade as the context window fills with stale tool output and dead ends — they call it context rot — and the fix is a six-command workflow loop (new → discuss → plan → execute → verify → ship) that keeps the main window clean while spawning fresh-context subagents for research, planning, and execution. The installer transforms a single source tree into native files for 15+ runtimes — Claude Code, OpenCode, Gemini CLI, Copilot, Cursor, Windsurf, et al. — and the project ships with a public security audit and explicit governance after the trust break with the upstream. Two patterns in one repo: the meta-prompting category is now mature enough to fork over governance, and "cross-runtime portability" has stopped being a feature and become a baseline expectation.

A community project from nexu-io that doesn't ship its own agent — it detects whichever coding-agent CLI you already have on your PATH (Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Hermes, Pi, Kimi, OpenClaw, Devin, Qwen, Mistral Vibe, Kiro, Kilo, DeepSeek TUI, Qoder) and uses it as the brain behind a design workbench: 132 composable skill files across 32 prototype modes and 9 deck modes, 150 portable design systems as plain DESIGN.md specs, turn-1 discovery forms that lock the brief, and sandboxed iframe previews with HTML/PDF/PPTX/MP4 export. The pattern is the point: a viral cloud-only Anthropic product (Claude Design, April) gets a credible local-first community alternative within a month, and the alternative is harness-agnostic by construction. The harness is the OS, the design system is the brief, the model is whichever one you trust this week.

A small Shell+Markdown skill pack with one of the better elevator pitches in the category — codified taste rules for design, front-end, and copy, drop them in your skills directory, and the agent stops defaulting to bland Tailwind boilerplate. Same pattern as Tufte-claude-skill last week (charts) and the OpenMobius/medharness packs (verticals): taste is now a packaging format. The interesting wrinkle here is that "good taste" is being articulated as negative constraints — what not to produce — which is closer to a refined style guide than a prompt template. The next month will tell us whether taste packs from named designers (Tognazzini, Bringhurst, Norman) start showing up as installable skills, the same way trading methodologies and legal practice areas did this quarter.

mvanhorn shipped a recency-aware research skill that queries a fixed set of high-signal sources — Reddit, X, YouTube, Hacker News, Polymarket, Bluesky, TikTok, Instagram — over a configurable trailing window and synthesises a grounded summary back into the agent's context. The pitch is the dataset cocktail: each source is plumbed through its own connector with a stated recency floor, so the synthesis is provably not just whatever the model remembers from training. It's the same skill-packaging pattern as the trading and legal packs, but pointed at the harder problem under "deep research" — recency. Expect adversarial research skills (auto-debate, cross-source contradiction, evidence-grading) to land in the same lane over the next few weeks; this one's first.

safishamsi's coding-assistant skill (cross-harness from day one — Claude Code, Codex, Cursor, Gemini CLI, OpenCode) does one specific thing well: ingest a folder of heterogenous content — application code, SQL schemas, R scripts, shell scripts, docs, papers, images, video — and emit a single queryable graph, stitched with tree-sitter parsers, GraphRAG retrieval and Leiden community detection. The pitch is the unified representation: app code, database schema, and infrastructure end up in one graph the agent can traverse, instead of three separate retrieval pipelines that the model has to triangulate across. Where pluck went small (98% fewer read tokens) and Semble went token- lean for code search, graphify goes broad — every shape of project artefact under one query language.