MetaWatch Patch 2026.07

The meta,
as of today.

Patch notes for agentic development - models, harnesses, and the config that matters. Updated 25 Jul 2026
▼ Nerf Gemini loses the vision crown - led no tracker; Fable 5 takes it ▲ Buff Sol & Terra sweep the value slots - subagent, delivery, serving podiums ▼ Nerf Sonnet 5 drops off the board - outclassed on performance:cost ● New Explainers land - parameters, the tokenizer tax, benchmarks

Models

by job
Overall ■ HOLD Claude Fable 5 2 Claude Opus 5 3 GPT-5.6-Sol

Big-model smell - overall intelligence, judgement, and taste. The vibes category this site exists to argue.

Orchestration ■ HOLD Claude Opus 5 2 Claude Fable 5 3 GPT-5.6-Sol

Decomposing briefs, dispatching subagents, and owning the judgement calls at the top of the stack.

Subagent ▲ RISING Grok 4.5 2 GPT-5.6-Sol 3 GPT-5.6-Terra

Fast, cheap, and RL'd to write code with maximum efficiency. Just don't set them loose on their own.

Delivery engine ■ HOLD GPT-5.6-Sol 2 GPT-5.6-Terra 3 Claude Opus 5

Acceptance criteria in, working code out. The slot where relentless beats tasteful, every time.

Diff review ■ HOLD GPT-5.6-Sol 2 Claude Opus 5 3 Claude Fable 5

The second pair of eyes on every diff. Recall over politeness, and never the model that wrote it.

Plan review ■ HOLD Claude Fable 5 2 Claude Opus 5 3 GPT-5.6-Sol

Judging proposals before any code exists - business, UX, and maintenance implications included.

Frontend - No Reference ● NEW Kimi K3 2 Claude Opus 5 3 Claude Fable 5

Starting from a loose brief and making the visual decisions. Taste, range, and restraint matter more than raw coding strength.

Frontend - Reference ● NEW GPT-5.6-Sol 2 Gemini-3.6-Flash 3 Kimi K3

Turning an existing interface into working code. Visual reasoning, fidelity, and implementation quality decide the podium.

Writing & marketing ■ HOLD Claude Fable 5 2 Claude Opus 5 3 Gemini 3.1 Pro

Words a human will read - copy, positioning, long-form. A very different ranking from the code slots.

Harnesses

By who you are
Most people ■ HOLD Codex 2 Cursor 3 Claude Code

The one tool you'd roll out across the whole skill spectrum, from power users to complete beginners.

Developers ■ HOLD Claude Code 2 Codex 3 Cursor

For terminal-comfortable devs who want the strongest defaults without living in their config files.

Power users ◆ NICHE Pi 2 Codex 3 Claude Code

Substrate, not defaults - maximum extensibility for people with opinions about every layer of the stack.

Cloud agents ● NEW Cursor 2 Devin 3 Claude Code

Full coding environments away from your laptop. The environment, verification loop, and handoff matter as much as the model.

Vibe coding ● NEW Lovable 2 Replit 3 v0

Idea to deployed app for people who do not write code. Ease of use wins only when the platform also protects the result from slop.

Serving

by use case
Chatbots ■ HOLD DeepSeek V4 Flash 2 Gemini 3.5 Flash-Lite 3 DeepSeek V4

High-volume conversation where unit economics decide everything. Good enough beats brilliant here.

Cost-conscious agents ■ HOLD Grok 4.5 2 GPT-5.6-Terra 3 GLM-5.2

Real agentic work inside a product that has to keep its margins. Performance per pound wins here.

High-intelligence agents ● NEW GPT-5.6-Sol 2 Kimi K3 3 Grok 4.5

When your product sells intelligence, buy the most capability per pound the frontier will part with.

Primitives

The config that matters
AGENTS.md Guide →

Less is more. If yours says “don't commit secrets”, you're behind the curve.

 Skills Guide →

Codify observed failure modes, not hypotheticals. Tight language, explicit intent.

 Hooks Guide →

Deterministic actions on harness events. Reach for one to kill a specific frustration.

Explainers

The concepts behind the picks the key concepts
Parameters, and big-model smell Guide →

What parameter counts actually are, why bigger models feel smarter, and the house theory behind our Overall crown.

 The tokenizer tax Guide →

Why price-per-token is a lie, why cost-per-task is the only honest number, and how to read any pricing announcement.

 The benchmarks worth believing Guide →

Three benchmark demolitions in sixty days, the trust hierarchy that survived, and how MetaWatch reads a leaderboard.

Featured

The shortlist
■ mattpocock/skills The skills collection for engineers. github.com ■ emilkowalski/skills The skills collection for designers. github.com ■ coreyhaines31/marketingskills Skills for marketers and sales. github.com ■ Agentic index Performance-per-dollar, indexed and current. artificialanalysis.ai ■ Text arena Writing ability at a glance. arena.ai ■ Design arena Design ability at a glance. designarena.ai ■ system_prompts_leaks Read a harness's mind. github.com

Our Meta

Public setup

An opinionated setup, stripped down to the public parts worth stealing.

Global instructions Start here Build yours

Build a lean AGENTS.md shared across your coding agents before adding skills.

MetaWatch owned 4 MetaWatch Collection

Our workflow bundles and focused portable skills.

Bundled third-party 3 Attributed sources

Licensed snapshots and adaptations, credited to their original creators. Inclusion is not endorsement.

Matt Pocock skill collection Complete attributed bundle Seven engineering and productivity skills that install together. Source: Matt Pocock codebase-designdomain-modelingimprove-codebase-architecturetddwayfindergrilling writing-great-skills The meta-skill for codifying repeatable knowledge into well-structured skills.
Marketing package Complete attributed bundle Six marketing workflows under one progressively disclosed router. Source: Corey Haines AI SEOcold emailcopywritingpricing strategyproduct marketing contextprogrammatic SEO
External prerequisite 1 Attributed source

Portable guidance that still depends on tooling outside MetaWatch Collection.

context7-cli Skill root + external CLI Current library documentation through Context7. Source: Upstash Prerequisite: Requires a current Node.js runtime, npx, and the external ctx7 CLI; the skill does not install them.
Paste into your coding agent 
Inspect the public MetaWatch Collection repository and its registry:
https://github.com/Hxnnyy/MetaWatch-Collection

Review the target project's existing agent instructions and harness configuration. Select the smallest useful complete registered bundle root from the current registry. Treat that root as an indivisible source tree so its relative references and entrypoints remain intact. Do not install a nested workflow skill by itself.

For third-party material, read the bundle README and provenance file, preserve provenance and the referenced licence, and distinguish exact snapshots from adaptations. If context7-cli is selected, confirm the external Node.js, npx, and ctx7 CLI prerequisites rather than claiming the skill installs them.

Show me the proposed files and changes before applying them, explain why each choice fits, and wait for approval.

When approved, merge rather than overwrite existing project instructions and preserve local customisations. Do not invent an installation location; use one documented by the target harness or selected by me.

Install only public, safe, portable material. Exclude credentials, private context, account-specific data, machine-specific paths or settings, private integrations, and destructive hooks. After applying the approved changes, run the existing checks, inspect the diff for portability hazards, and summarise exactly what changed.
Browse MetaWatch Collection

Copied - paste it to your agent