Skip to main content
GatiFlowIntelligence Platform
Deep DiveLoginRegister

GatiFlow Intelligence Platform

TermsPrivacyComplianceOpt-outMethodologyAcademyChangelogStatus

← All Deep Dives
Saturday Deep Dive

Jane Street Opens an Internal Agent Skill Just as the Format Gets a Neutral Home

Saturday edition · November 21, 2026 · 1760 words · 9 min read

On November 3, 2026, Jane Street, the quantitative trading firm that runs more than 30 million lines of OCaml, created a public GitHub repository holding a single set of instructions for AI agents. By November 19 it had 10,175 stars. The count is the visible part; the README carries the substance. Jane Street says the skills in the repository "are exported from our internal repository, where we use them day to day." Jane Street has published working agent know-how in the same open format that, twelve days earlier, moved under neutral governance. A real internal artifact, shipped in a portable standard, is the signal this piece follows.

GatiFlow's GitHub collector first logged janestreet/agent_skills on November 13 at 05:50 UTC with 7,895 stars and kept it in the tracked set for 13 collection cycles across four calendar days. At the last of those readings, on November 16 at 23:51 UTC, the repository stood at 9,577 stars: 1,682 more, a rise GatiFlow's velocity model recorded as 21.3 percent against the November 13 baseline.

The pace did not fade inside that window; it picked up. Measured between the 05:50 UTC readings, the repository added 451 stars on the first day and 335 on the second, then 457 on the third. The last full day of tracking, from 23:50 UTC on November 15 to 23:51 UTC on November 16, added 612, and each of the four six-hour readings on November 16 rose between 1.1 and 1.9 percent. Forks went from 513 to 643 over the same span. The owner grew too: the Jane Street organization account moved from 3,486 to 3,693 followers in GatiFlow's data, a 5.9 percent gain in under four days.

Public data on either side of that window tells the same story. A third-party tracker logged 2,593 stars on November 10, three days before GatiFlow's first reading, and the GitHub page showed 10,175 stars and 694 forks on November 19, after the last tracked cycle. The repository has gained stars at every public or proprietary reading we have.

That persistence is why the theme qualified for this piece. GatiFlow publishes a trend signal only once two independent sources confirm it, or one source repeats across two collection cycles. GitHub reported janestreet/agent_skills in 13 cycles over four days, well past that bar, and the repository belongs to an established organization, which GatiFlow's theme selection ranks above personal projects when it picks a Deep Dive.

The content is small by design. The repository holds one skill, socratic-tutor, written in the SKILL.md format and released under the MIT license. The README says the skill "turns an LLM into a strict personal tutor." It places the learner with an opening diagnostic, tracks progress on a six-level skill map, asks one question at a time, schedules spaced review and closes with a final assessment.

The skill is plain-language instructions in Markdown, and installing it means copying its folder into an agent's skills directory. Its craft shows in the details. It opens with a guard that stops the agent from launching a tutoring session when the user only asked a one-off question, and explains the trade-off in one line: "A false start costs the user an unwanted quiz; a confirmation costs one word." Builders noticed. A Hacker News post titled "Socratic Tutor," linking straight to the skill file, collected 722 points and 171 comments after it was submitted on November 10.

Jane Street also sets clear terms for the release. It is "not looking for new skills or for pull requests that change the text of existing ones," and welcomes issues when something does not work as described. The source of truth stays inside the firm; the public repository is a copy. That is a deliberate model for publishing internal know-how without turning it into a community project to maintain.

On October 22, 2026, Anthropic announced it was donating Agent Skills, which it describes as "the open standard that gives AI agents specialized knowledge and workflows," to the Agentic AI Foundation (AAIF) under the Linux Foundation. Anthropic's announcement says "More than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have integrated Agent Skills" since the standard was published in December 2025, with 1 million skills published and nearly 280 million installed.

The donation was one of four projects AAIF added at AGNTCon + MCPCon in San Jose, alongside OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox. That took the foundation to ten projects, a roster that also includes MCP, goose, AGENTS.md, A2A, agentgateway and Agent Router. At the same event, MCP lead maintainer Den Delimarsky reported cumulative MCP SDK downloads of 4.17 billion, up from about 3 billion a month earlier.

Read together, the two events explain each other. Jane Street's README tells users to drop the skill into "a coding agent that supports the common SKILL.md format," or to paste it into a plain chatbot. One file, written once, runs across the platforms that adopted the standard. That portability is what a neutral home is meant to protect: a company can invest in writing down how it works without betting on a single agent vendor. Jane Street's release puts that investment in public view.

For product and platform teams, the practical question is where a skill fits next to MCP. Atlan's decision guide puts it in one line: "Skills answer how, MCP answers access, and most production agents need both." Its sorting test is whether the thing an agent needs changes between invocations, which points to MCP, or stays stable for weeks, which points to a skill. A guide in the awesome-agents-on-google-cloud project gives the cost argument: "Start with a skill when what's missing is know-how; a Markdown file is the cheapest thing to try." It adds: "Reach for MCP when you need a secure way to give the agent access to the outside world."

Jane Street's tutor is the pure case of the first layer. It needs no live data, no credentials and no connection to any system; it encodes a teaching procedure. That makes it a useful template for teams sorting their own backlog: style guides, review checklists, onboarding drills and incident playbooks are procedure, and a skill is the cheap first step; anything that reads a database or calls an API is access, and belongs behind MCP.

The choice of a tutor fits how Jane Street hires. According to Java Code Geeks, citing the firm's interviewing page, "most of the software engineers it hires arrive with no OCaml or functional programming experience at all." Coding interviews "are done in whatever language the candidate already knows best," new engineering hires "go through months of internal training and team rotations," and a dedicated four-week OCaml Bootcamp brings incoming traders and non-engineering staff up to speed. A tutor that teaches by asking questions is consistent with a firm that hires for ability and builds fluency in-house. The README does not say which teams use the tutor, so this is a fit with the firm's stated culture, not a confirmed use case.

For recruiters, the signal is about engineering culture rather than headcount. Jane Street did not appear among the companies in the hiring signals GatiFlow tracked between November 13 and 19, so the release reads as a firm showing how it works, not as a recruiting push tied to agent tooling. GatiFlow will report it if that changes.

For investors, the opportunity sits in the layer around the files more than in the files themselves. Anthropic's announcement sets out a roadmap that includes versioned releases, a formal process for proposals, and security guidance for skill authors and the people who install them. Versioned releases and security guidance are still roadmap items for a format already integrated by more than 70 platforms. Companies that make skills safe to version, review and share inside an enterprise are building on a standard with neutral governance and wide reach.

For developer-tool companies, the cost of reaching agent users has dropped to a Markdown file. A skill that teaches an agent how to use your SDK or API well travels to every platform that reads the format, and Jane Street's release shows that even a single, well-written skill can earn wide attention from developers.

For product and platform teams, two habits follow. First, treat third-party skills like dependencies: a skill is instructions an agent will act on, so it deserves the same review a new package gets, at least until the standard's security guidance is published. Second, consider Jane Street's publishing model if you want to share internal know-how: keep the source of truth internal, export a copy, and state plainly what contributions you accept.

The AAIF's MCP Dev Summit Nairobi runs November 19 to 20, 2026; any update to the skills roadmap announced there is the next milestone for teams planning around the format. GatiFlow will keep tracking two things: whether a second independent source, such as a package registry, research output or hiring data, confirms the theme, and whether Jane Street adds more skills to a repository its README already calls "a collection."

A firm that teaches its engineers OCaml in-house has published a tutoring skill it says it uses day to day, in a format now governed by a neutral foundation. The star count shows developers noticed; the README shows why it matters.

Coverage note: this theme was seen by one source (GitHub) on 4 of the last 7 days and has not yet been confirmed by a second independent source.

Sources:

- janestreet/agent_skills (README) - GitHub (https://github.com/janestreet/agent_skills)

- socratic-tutor/SKILL.md - janestreet/agent_skills - GitHub (https://github.com/janestreet/agent_skills/blob/master/socratic-tutor/SKILL.md)

- Socratic Tutor - Hacker News (https://news.ycombinator.com/item?id=50438261)

- Jane Street Agent Skills - aiskills.one (https://aiskills.one/skill/janestreet-agent-skills/)

- Donating Agent Skills to the Agentic AI Foundation \ Anthropic (https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation)

- Agent Skills gets a neutral home, and your organisation's know-how becomes portable - i-SCOOP (https://www.i-scoop.eu/agent-skills-agentic-ai-foundation/)

- AGNTCon + MCPCon 2026: Agentic AI Foundation Adds Four Projects as MCP Turns to Security - daily.dev (https://daily.dev/posts/agntcon-mcpcon-2026-agentic-ai-foundation-adds-four-projects-as-mcp-turns-to-security-fwqfyi00y)

- Agent Skills vs MCP: Architecture and Decision Guide [2026] - Atlan (https://atlan.com/know/ai-agent/ai-agent-skills/agent-skills-vs-mcp/)

- Skills vs MCP: when to use which (https://ykdojo.github.io/awesome-agents-on-google-cloud/skills-vs-mcp.html)

- OCaml Quietly Running the World's Financial Infrastructure - Java Code Geeks (https://www.javacodegeeks.com/2026/09/ocaml-quietly-running-the-worlds-financial-infrastructure.html)

- MCP Dev Summit Nairobi - Linux Foundation Events (https://events.linuxfoundation.org/mcp-dev-summit-nairobi/)

Disclaimer: This article is generated by GatiFlow Intelligence for informational purposes only. It does not constitute investment advice, recruitment recommendations, or legal guidance. All data is derived from public sources and AI analysis — verify independently before making decisions. Past trends do not guarantee future results.

Where this came from

Every Deep Dive starts from GatiFlow's own pipeline: 13 public developer sources, collected every six hours, with a confidence score and the evidence behind each signal. The same signals, filtered to the topics you follow, are a JSON API.

No credit card required.

Get the next one by email

One article every Saturday morning in your time zone. No account needed, and nothing else is sent to the address.

Double opt-in: you confirm by email first. What we store, and for how long, is in the privacy policy.

Tell me I am wrong

Corrections, the version of this you have lived through, or what you would like covered next. It reaches me directly and is never published. It is kept for two years so it can be read and answered; the privacy policy has the details.

0/2000