Thursday, August 20, 2026

A hot Runpod skill for moving workloads between environments. It targets migration work that appears when GPU capacity, pricing, or a provider relationship changes.

Useful if your agent fleet needs an exit path rather than a pile of provider-specific assumptions.

Skills.sh · infrastructure · medium

A current skill for turning product requirements into interface work. It gives coding agents a sharper brief before they start laying out screens.

Good UI work starts with hierarchy and flow decisions, not a prettier first draft.

Skills.sh · design · medium

A hot companion skill that reviews interface output for generic low-signal patterns. Use it after an agent produces a landing page, dashboard, or internal tool.

A cheap review loop can stop fast agent output from becoming another bland, hard-to-use surface.

Skills.sh · quality · medium

A narrowly scoped Dart skill for adopting primary constructors where they fit. Small language rules can keep coding agents aligned with an evolving codebase style.

Boring constraints can prevent a hundred tiny review comments.

Skills.sh · coding · low

A hot workflow skill from Garry Tan’s gstack project. It packages repeatable product and engineering work instead of treating every prompt as a fresh start.

Worth dissecting for patterns you can adapt to client-agent operating procedures.

Skills.sh · workflow · medium

A current skill for auditing a website’s messaging, structure, and obvious issues before a sales or growth workflow hands traffic to it.

For Vesara-style outbound, the landing page is part of the qualification loop.

Skills.sh · growth · medium

santifer/career-ops

An open-source AI job-search workflow that scans listings, scores fit, tailors a CV, and tracks applications from a local coding CLI. It is a concrete vertical-agent pattern.

MIT · 198 stars today · 65,980 total

AgriciDaniel/claude-ads EARLY

A Claude-first paid-media operations skill spanning twelve ad platforms, with source-grounded audits and structured reports. It is relevant to bounded agent workflows around revenue operations.

MIT · 51 stars today · 8,273 total

eneskirca/nodeterm EARLY

A node-based terminal manager for parallel AI coding sessions, backed by tmux. It makes concurrent agent work visible instead of burying it across terminal tabs.

MIT · 78 stars today · 966 total

letta-ai/letta-code EARLY

A terminal-oriented coding agent from Letta that centers persistent memory and identity. Test it against stateless coding loops on projects that span many sessions.

Apache-2.0 · 16 stars today · 3,063 total

abhigyanpatwari/GitNexus

A browser-based code intelligence tool that builds an interactive knowledge graph from a repository or archive without a server. Useful for orienting agents in unfamiliar code.

MIT · 41 stars today · 45,601 total

youssofal/MTPLX EARLY

An Apple Silicon inference project claiming faster MLX throughput through native speculative decoding. It is a local-inference benchmark candidate, not a result to take on faith.

Apache-2.0 · 43 stars today · 1,460 total

OpenRouter says it is joining Stripe. The deal puts a model-routing layer next to payments and billing, where per-request economics and provider choice become product controls.

HN

A popular Claude Code request asks for AGENTS.md support. Teams want durable repository instructions that survive individual prompts and contributors.

HN

Unsloth published Dynamic 3.0 GGUF guidance for local model quantization. Test whether its quality, memory, and latency trade makes a smaller deployment viable.

HN

Simon Willison tests smolmachines and smolvm for running untrusted Python and JavaScript. The important question is whether the isolation boundaries match agent tool permissions.

Simon Willison

Every examines Codex economics for an engineering team. Code can get cheaper while feedback loops, product judgment, and deployment bottlenecks stay expensive.

Every

A community thread discusses Project Parka, which sits through meetings and assigns follow-up work to Claude agents. The hard part is giving agents the right authority and review loop.

R/CLAUDEAI

One builder wired Claude Cowork into an iPhone widget showing mission status, progress, changed files, and the next step. Ambient visibility may matter more than another chat interface.

R/CLAUDEAI

A thread asks whether teams can list each connected MCP server, its owner, and what it can reach. If the answer is no, access review and incident response will be guesswork.

R/MCP

A builder shares an MCP server that deploys an app from a single tool call and returns a private URL. Production use still needs clear secret handling and approval boundaries.

R/MCP

A founder gave an autonomous agent a domain and a constrained budget to build a project over repeated sessions. Persistence changes both capability and the shape of oversight.

R/AI_AGENTS

AI-native accounting startup Rillet raised a $100 million Series C at a $1 billion valuation, led by Iconiq, after saying it doubled ARR over the past three months.

$100M · Series C

AI infrastructure startup Velatir raised €5 million to accelerate adoption across Europe, according to Sifted.

€5M · Seed

A harness-scaling paper reports 95.3% raw accuracy on Terminal-Bench 2.1 for a stated $15 frontier-model run. It studies execution systems around long-horizon agents.

420 upvotes upvotes · Aug 15

This paper asks why skills improve agent performance in some cases and fail in others. It fits teams turning prompt lore into reusable operating procedures.

149 upvotes upvotes · Aug 14

A method for fine-tuning long-horizon agents with minimal GPU requirements. It targets the cost and branching problems that make agent training impractical.

96 upvotes upvotes · Aug 18

An edge-native MoE serving system that adapts execution to bandwidth on personal machines. Relevant if local inference economics matter in deployments.

64 upvotes upvotes · Aug 17