Today I spoke to the AI Transformation team at a Fortune 500 financial services firm. They wanted to know how many customers use Coder for skills distrubtion. The honest answer: Not many. I dug in a bit.
They have 19,000 skills, running across 175 unique types of Coder workspaces.
That’s two unique, codified procedures per employee. Not prompts. Not one-off scripts. Skills, meaning structured, reusable, version-controlled instructions that any employee’s agent can load on demand. They’re capturing how their business actually runs, writing it down as markdown, and shipping it straight into the AI tooling every employee already opens every day. Workspaces aren’t just the sandbox where code gets written anymore. They’re the distribution channel for the company’s own institutional knowledge.
That is workflows-as-code, in production, at a scale I hadn’t heard anyone quantify yet. It’s also the clearest evidence yet for something I’ve been circling in this blog for months.
Knowledge Is Table Stakes Now
Here’s how I think about it. AI is knowledge-as-code. It’s the world’s information distilled and made instantly accessible. That’s genuinely remarkable, and also now table stakes. Every business has access to roughly the same base intelligence, whether they’re pulling from frontier labs, open source models, or something tuned and run on-premises.
The lag between a frontier model releasing something remarkable and an open or fast-follower lab shipping the functional equivalent has compressed from months to weeks. That’s not a knock on any one lab. It’s what commoditization looks like in real time, and it’s the same cloud-native pattern I heard echoed by a whole roomful of executives: the infrastructure gets commoditized first, and the winners are the ones who redesign their operating model around it.
What is not commoditized is your workflow: the specific sequence of decisions, judgments, and steps your team follows to deliver value to a customer. That remains proprietary. If you’re a knowledge worker, it’s your economic engine. I’ve written before about the gap between the AI-haves and the AI-have-nots. The financial services team I talked to is a data point for where that gap actually goes: you move from knowledge-as-code to skills-as-code.
The Skill File Is Quietly One of the Most Important Ideas in AI Right Now
The emergence of skills files and markdown-based instruction sets is revelatory, and it’s not just one company’s idea. Anthropic’s Agent Skills format, packaged as a SKILL.md file, is now an independent open standard with a public spec and reference SDK, and it’s already spread well past Claude. Microsoft has adopted it inside VS Code and GitHub; Cursor, Goose, Amp, and OpenCode support it too, and OpenAI has quietly shipped structurally identical directories in ChatGPT and Codex. When four competing labs converge on the same file format within months of each other, that’s not a feature. That’s infrastructure.
I went and built this myself before I ever heard the 19,000-skill number. It’s a small-scale version of the same idea: I kept typing the same handoff paragraph into chat every time a session ran long, until I turned it into a reusable skill instead. Multiply that instinct by an enterprise, and 19,000 stops sounding like an outlier and starts sounding like the natural endpoint.
The ability to codify a workflow in a structured way and hand it to an agent as a repeatable set of tasks sounds almost too simple, but the implications are significant. You get the benefits of a non-deterministic technology, a model that can reason, adapt, and handle edge cases, combined with the value of it following a deterministic process. Best practices. Business rules. Dare I say your “alpha.” The stuff that took months for a new hire to internalize consistently, an agent now executes instantly.
Onboarding goes from a problem to an advantage. Getting someone to follow a process reliably, every time, not just when they remember or feel like it, is genuinely hard. It’s a huge part of what makes scaling a knowledge-work business difficult. An agent with a well-written skills file doesn’t have that problem. It follows the process to the letter, usually, at whatever speed and volume you need.
This isn’t about staff augmentation or displacement. That conversation is real but also a distraction. The more interesting conversation is asking: if you could run your core business processes ten times faster, what would you do with that? How many more customers could you serve? How many more deals could you close? What would you build that you currently can’t afford to?
The market is already pricing this in. Anthropic’s own skills directory now ships partner-built skills from Atlassian, Canva, Cloudflare, Figma, Notion, Ramp, Sentry, Stripe, and Zapier, and Fortune 500 companies are reportedly running skills in production not just for coding but across legal, finance, accounting, and data science. A venture firm put it bluntly: for the last two decades, developer-tools distribution meant getting an SDK installed one team at a time. Agent skills are becoming the new SDK, landing directly in the agent’s context instead of a dependency tree.
Govern Skills Like You Govern Code
I know, I know. AI can be dangerous and skills are yet new attack vector. Skills-as-code has a supply chain, and supply chains get poisoned. Snyk’s February 2026 ToxicSkills audit scanned nearly 4,000 publicly published skills and found that roughly 37% carried at least one security flaw, with 13.4% rated critical, including confirmed malware distribution and credential exfiltration. That’s not a reason to avoid skills. It’s the exact reason regular audits matter, and why I keep coming back to the idea that vibe coding at enterprise scale is inseparable from a governance mindset.
The financial services team’s 19,000 skills are internally authored and centrally distributed through workspaces, which is precisely the model that keeps this safe. They treat the skills development lifecycle like the software development lifecycle. Skills as governed, version-controlled, audited and run through a pipeline. They control the infrastructure, and it’s not something an employee installs off a public registry on a whim. Anthropic clearly sees the same risk. Its enterprise plans added organization-wide skill management and audit trails specifically so admins can control which skills are provisioned company-wide, rather than leaving it to individual chat windows. The pattern holds: the more valuable skills-as-code becomes, the more it needs the same discipline you’d apply to any other production code path.
Memory as Code Is the Next Layer
There’s a second piece I’ll only touch on lightly. Maybe more in a future post.
When agents execute workflows, they generate something we haven’t had before at scale: a full log of the thinking. Every prompt, every chat line, every inference can be logged. Every intermediate step and decision made along the way can be logged. The intent from the human, or the agent, that initiated the task can be logged. It can be stored, harvested, analyzed. This is the same durability problem I’ve been chasing on the chat side: if you’re throwing away context, you’re throwing away competitive advantage.
The inference layer is a source of business intelligence. You can scan thousands of workflow executions and start to see patterns. Where is the process breaking down? Where is it surprisingly efficient? What are agents doing that your human team wasn’t doing, or vice versa?
Then you can take those insights and write new skills files. Refine the process. Feed it back in. The loop closes on itself. At 19,000 skills and counting, that feedback loop isn’t hypothetical anymore. It’s a live operating model.
What you end up with is not just AI-assisted work. It’s a system where workflows generate knowledge, knowledge improves workflows, and the intelligence layer underneath is a commodity you can swap or upgrade as the market evolves. Frontier model versus open source, cloud versus on-prem, self-hosted versus API-metered: it doesn’t matter. The proprietary value sits in the infrastructure, as code.
We’re at a genuinely strange moment for businesses whose entire economic model is built on knowledge work. The tools to codify and execute those workflows at scale exist right now, and at least one Fortune 500 team is already running 19,000 of them through production infrastructure. The businesses that make that their operating model are going to outpace everyone else.
If you could write down your most valuable business process as a set of instructions an agent could follow, what would be the hardest part to capture?
By the Numbers
- 19,000 skills — built by one Fortune 500 financial services AI Transformation team in six months
- 175 unique workspace types — the range those skills run across, all inside Coder
- 6 months — the time it took to go from zero to that scale
- ~4,000 skills — the size of Snyk’s February 2026 public-registry audit
- 37% — of those audited skills carried at least one security flaw, 13.4% rated critical
- 9 companies — Atlassian, Canva, Cloudflare, Figma, Notion, Ramp, Sentry, Stripe, and Zapier, all shipping partner-built skills in Anthropic’s own directory
- 4 major agent ecosystems — Claude, Microsoft VS Code/GitHub, Cursor/Goose/Amp/OpenCode, and OpenAI ChatGPT/Codex, now converging on the same
SKILL.mdformat - 0 — the number of those 19,000 skills that had to be re-explained to an agent a second time, once written down