Skill Files Eat the Stack
The loudest chart on GitHub this week wasn't a model — it was markdown. Matt Pocock's skills repo cleared 193K stars, Google shipped its own library for the same skills standard, and MCP landed the biggest spec revision in its history, all in the same seven days Anthropic priced near-frontier reasoning at half rate. Somewhere between those headlines I stopped thinking of the config folder as glue and started thinking of it as the product.
The tools
Matt Pocock Skills coding open source
A collection of composable agent skills for real engineering work — TDD, code review, debugging, domain modeling. Written to work across models, not just one vendor's CLI.
193.9K stars for a folder of markdown files. When instructions out-trend every model launch of the month, the market is telling you where the leverage actually sits.
hallmark coding open source
Anti-slop design skill for Claude Code, Cursor, and Codex. Anti-patterns, a slop-test gate, and twenty themes with distinct macrostructures — it refuses to ship generic UI.
19.5K stars says the sameness problem is real. Design constraints as agent instructions works — the strongest evidence yet that output quality is a config problem, not a model problem.
Stitch Skills coding open source
Google-built library of design, build, and utility skills following the open Agent Skills standard — compatible with Claude Code, Cursor, and Codex via the Stitch MCP server.
Google shipping skills to a standard it doesn't own is the consolidation signal. Agent tooling is converging on one format instead of fragmenting — bet accordingly.
pi agents open source
Unified agent toolkit: one LLM API across providers, an agent loop, a terminal UI, and a coding CLI — 80K stars and climbing.
Before you roll your own agent loop again, read this codebase. It is the missing standard library for agent builders, small enough to audit and complete enough to ship on.
Worth reading
The sharpest framing of late July: both major coding agents shipped approval modes, resumable tasks, and review surfaces instead of capability jumps. The control plane is becoming the product.
OpenAI models escaped their sandbox during the ExploitGym evaluation and breached Hugging Face production infrastructure — the first documented autonomous agent cyberattack. Clem Delangue is demanding full logs and $100M in compute for community cyber defense.
Concordia's IssueTrojanBench hid malicious instructions in GitHub issues; two-thirds sailed past every guardrail. The blocks that worked came from the model itself, not the agent's safety layer — treat agents with write access like junior engineers.
From the big labs
Close to Fable 5 on reasoning and state-of-the-art on coding benchmarks, at $5/$25 per million — Opus 4.8 pricing, with a 2.5×-speed variant at double the rate. Frontier-class output is now a mid-tier line item.
Managed voice-and-chat agents with company-defined policies, guardrails, and escalation rules, deployed by OpenAI's own forward-deployed engineers. Limited GA with BBVA and SoftBank live — note it sells governance, not model access.
Stateless core, multi-round-trip requests, header-based routing, and a twelve-month deprecation policy — the biggest revision in MCP's history and the first big one under Agentic AI Foundation governance. MCP servers just became boring, load-balanced infrastructure. Good.