Agents That Show Their Work, Day 251
October 21, 2026. The newcomers on today’s board have one thing in common: they don’t just ask you to trust what an agent produces, they make it checkable. Cloudflare’s security audit skill verifies every finding a second time before reporting it. graphify builds a map of your codebase where every connection comes with an explanation. Around them, the plugin layer keeps splitting in three directions: official plugins from the vendors, marketplaces that work across vendors, and libraries of ready-made agent personas.
cloudflare/security-audit-skill: An Audit You Can Verify
What it is. A skill for coding agents that runs a security audit in several phases, then independently checks each finding before reporting it. The results come out in a machine-readable format instead of a prose report.
The problem. AI security reviews have a credibility problem. Ask an agent to find vulnerabilities and it will happily find some, including ones that do not exist. A security team that has to manually re-check fifty findings to discover forty are false alarms will stop using the tool within a week.
Who was there before. Traditional static analysis tools such as Semgrep, CodeQL, Snyk and SonarQube follow predefined rules: precise, but blind to anything outside their rulebook. On the AI side, Claude Code ships its own security review command, and GitHub offers AI-generated fixes for code scanning alerts. Those are good at spotting issues but leave verification to the human.
Is it new? The idea of an AI security review is not. The two design choices are what matter:
- Independent verification. A separate pass tries to confirm each finding before it is reported. That trades speed for trust, which is the right trade for security.
- Machine-readable output. Findings that a program can parse can flow straight into a CI pipeline, a ticket tracker or a compliance dashboard, instead of living in a chat window.
Together, those turn an AI audit from “a second opinion you have to double-check” into something closer to a tool a security team can put into a pipeline. Coming from Cloudflare, a company whose business is security at internet scale, it also carries a credibility most community skills cannot.
Graphify-Labs/graphify: A Map Where Every Line Is Explained
What it is. A skill for Claude Code, Cursor, Codex and Gemini CLI that turns a codebase, along with its documentation, database schemas, config files and PDFs, into a knowledge graph an agent can query. It parses code locally and deterministically, explains every connection it draws, and uses no vector store.
The problem. Agents waste a lot of effort figuring out how a codebase fits together. They grep, open files, follow imports and often guess. Most of that understanding is lost at the end of the session.
Who was there before. Sourcegraph built its business on code intelligence for humans. Aider pioneered the “repo map,” a compact summary of a codebase built with tree-sitter. Cursor and similar editors index code into vector embeddings and search by similarity. codegraph, which trended on Day 231, also builds a pre-indexed graph for agents.
Is it new? It extends existing work in two meaningful ways:
- Beyond code. Most code tools stop at source files. graphify pulls in SQL schemas, configs, docs and PDFs, which is where a lot of a system’s real meaning lives, especially in enterprise codebases.
- Deterministic and explained, no vectors. Vector search returns “things that look similar” with no reason attached. A graph built from parsing returns “this function writes to this table, here is the line.” For regulated environments where you need to justify why an agent touched something, that difference matters.
It also fits a pattern this series has noted before with PageIndex: a move away from vector search toward structured, explainable retrieval.
DietrichGebert/ponytail: Teaching Agents Restraint
What it is. A skill that tells an agent to behave like “the laziest senior dev in the room.” Its whole philosophy is that the best code is code you never wrote.
The problem. Coding agents over-build. Ask for a small fix and you may get a new abstraction layer, three helper files and a config option nobody asked for. Every extra line is something a human has to review and maintain.
Who was there before. Engineering culture has said this for decades with principles like YAGNI (“you aren’t gonna need it”) and “keep it simple.” Many developers already paste similar rules into their own CLAUDE.md files. Larger skill collections such as superpowers include guidance on scope discipline.
Is it new? Technically, no. It is a carefully written set of instructions. Its value is that it packages a piece of senior engineering taste into something installable. It reached this series on Day 231 and is still growing, which says a lot about where developers’ frustration lies: not with agents that cannot write code, but with agents that write too much of it.
The Plugin Layer Splits Three Ways
Yesterday’s article noted that the agent ecosystem now has app stores. Today shows the different shapes those stores are taking.
Official, and beyond coding. anthropics/knowledge-work-plugins is Anthropic’s open collection of plugins aimed at knowledge workers rather than developers: people in finance, legal, sales and operations using Claude’s desktop agent for everyday work. The competitors here are Microsoft’s Copilot agents and OpenAI’s GPTs, which are tied to their own platforms. What is notable is the audience. The plugin model that started with coding agents is now being offered to people who never open a terminal.
Cross-vendor marketplace. wshobson/agents is a plugin marketplace that works across Claude Code, Codex, Cursor, OpenCode, GitHub Copilot, Google Antigravity and Pi. It is the opposite bet from the official directories: instead of one vendor curating for its own tool, a community curates once and ships everywhere. It fits the vendor-neutral thread running through orca and ai-memory.
Persona libraries. msitarzewski/agency-agents is a collection of specialized agent personas, from “frontend wizards” to “Reddit community ninjas,” each with a defined personality, process and deliverables. The idea of role-based agents is not new; CrewAI built a framework around it. agency-agents is not a framework, though. It is a library of well-written role definitions you drop into the agent you already use. Its value is curation, not invention.
Get Baozilla, Let's go!’s stories in your inbox
Join Medium for free to get updates from this writer.
Taken together, these show that “plugin” no longer means one thing. It can be an official extension, a portable package or simply a well-crafted prompt.
Returning Names
Several projects from the last two days are still on the board. Their stories have not changed, so here is the short version:
- stablyai/orca crossed 100K stars, the first vendor-neutral agent fleet manager to get there.
- MadsLorentzen/ai-job-search is still the top mover, a forkable job-search workflow running on Claude Code.
- headroomlabs-ai/headroom, the token compression proxy, and shiyu-coder/Kronos, the candlestick foundation model, both covered yesterday.
- ruvnet/ruflo and anomalyco/opencode continue as the swarm harness and the open-source coding agent.
The Mature Infrastructure
- vllm-project/vllm is the high-throughput engine many companies use to serve open models. Its key idea, PagedAttention, manages model memory the way an operating system manages RAM, which let it serve far more requests per GPU. Competitors include Hugging Face’s TGI, NVIDIA’s TensorRT-LLM and SGLang. It is no longer new; it is part of the plumbing.
- Comfy-Org/ComfyUI remains the standard node-based tool for building image and video generation pipelines locally.
What Today Actually Says
The newest projects on today’s board are less about what agents can do and more about whether you can trust what they did.
- Verified findings (security-audit-skill) instead of a list you have to double-check.
- Explained connections (graphify) instead of similarity scores.
- Less code (ponytail) instead of more to review.
That is the natural next stage. Once agents are capable, the bottleneck becomes the human who has to review their output. The tools that win from here are the ones that make that review faster, by being verifiable, explainable and restrained. For anyone deploying agents in a regulated company, these are the properties that turn an interesting demo into something an auditor will sign off on.
Baozilla Tech Daily
Key Statistics (October 21, 2026)
- cloudflare/security-audit-skill: 34,821 stars. Multi-phase AI security audit that verifies each finding and outputs machine-readable results.
- Graphify-Labs/graphify: 130,696 stars. Turns code, schemas, configs and docs into an explained knowledge graph without vector search.
- DietrichGebert/ponytail: 170,385 stars. Skill that pushes agents to write the minimum code needed.
- anthropics/knowledge-work-plugins: 30,929 stars. Anthropic’s open plugins for non-developer knowledge work.
- wshobson/agents: 41,370 stars. Plugin marketplace that works across seven different coding agents.
- msitarzewski/agency-agents: 162,633 stars. Library of specialized agent personas with defined roles and deliverables.
- stablyai/orca: 100,260 stars. Vendor-neutral workspace for running a fleet of coding agents.
Critical Signal: the newest agent tools are competing on trust, not capability.
- Cloudflare’s audit skill verifies every finding before reporting it, targeting the false-positive problem that holds back AI security reviews.
- graphify replaces vector similarity with a deterministic graph where every connection is explained.
- ponytail’s continued growth shows developers’ main complaint is agents writing too much code, not too little.
- Plugins now come in three shapes: official vendor extensions, cross-vendor marketplaces and persona libraries.
- Anthropic’s knowledge-work plugins take the extension model beyond developers for the first time.
Keywords for 20261021:
- Security-Audit
- Verifiable-Output
- Code-Graph
- Plugin-Marketplace
- Knowledge-Work
