Eight Open-Source AI Employees and the Portability Bet: Role-Based Agents You Own as Files

TL;DR

Reinventing.AI published eight role-based AI Employees on GitHub under the MIT licence on 19 September 2026: GTM Engineer, SEO/AEO, Web Dev, Social Media, Ad Manager, Sales, Customer Satisfaction and Chief of Staff.
Each employee is a folder of plain files — role description, operating contract, schedule, routines — that runs on the agent harness you already use, on your own machine, rather than a hosted product.
The repository’s own header counts 60 routines across Claude Code and 10 other harnesses; the launch press release says 59 routines on eleven harnesses and then names twelve. The real product is the portability claim, so those counts are exactly what a buyer should audit.
The safety posture is deliberately conservative: routines draft, fill and stage by default, and sending, publishing or spending only happen on channels the owner has explicitly released.

The AI agent industry has spent two years selling access. You rent a workspace, a seat, a per-task credit — and your automations live inside someone’s platform. On 19 September, a small company went the other way and published eight business roles as downloaded folders of text files, under a licence that permits commercial use, redistribution and resale.

What was actually released

Reinventing.AI, founded by Mark Fulton, released AI Employees as a public repository under the MIT licence on 19 September 2026. The eight roles cover the unglamorous operational surface of a small company: a GTM Engineer for launch positioning and outbound drafts; an SEO/AEO Employee producing one article per weekday plus indexing and rank review; a Web Dev Employee for site health and dependency review; a Social Media Employee drafting posts per platform with a veto window; an Ad Manager Employee that reads ad accounts and builds change lists; a Sales Employee running prospect sweeps and follow-ups; a Customer Satisfaction Employee sweeping the inbox with churn flags; and a Chief of Staff that reads every other employee’s run log and reports what quietly stopped. (Source : Reinventing.AI — Eight Open Source AI Employees on GitHub Under MIT License)

Mechanically, there is nothing exotic. An AI Employee is a folder containing a role description, an operating contract, a schedule and a set of routines. Installation is either a ZIP download or one command — npx ai-employees hire gtm-engineer --to <folder> — followed by pointing an agent at the folder and telling it to install the role. The agent researches the business from its website, builds a dashboard, schedules its own routines, and writes a morning brief describing what ran and what changed. (Source : GitHub — markfulton/ai-employees)

The routine distribution is where the work sits. Six of the eight roles carry seven or eight routines each; the repository’s per-role table totals 60. The launch press release says fifty-nine routines, and the repository’s own header line says “8 scheduled business roles, 60 routines, on Claude Code and 10 other harnesses”. (Source : GitHub — markfulton/ai-employees)

Roles as files beats roles as SaaS, until it doesn’t

The strategic argument for file-based roles is straightforward. A role contract written as markdown and scheduled by the harness is inspectable, diffable, forkable and version-controlled. When the agent does something wrong on Tuesday, you read the contract that produced it, patch it, and commit — instead of filing a support ticket and waiting for a vendor’s prompt update. Because everything is local, the employee reads from the same logged-in browser session you use and drives it the way a person does, which sidesteps an entire class of API-access negotiations.

The trade is equally straightforward, and the marketing is quiet about it. Files are portable; state is not. Browser profiles, credentials, session cookies, account permissions, accumulated context and the operator’s tolerance for a routine that misfires at 7am all stay with you. Portability across harnesses means the instruction layer moves, not the environment. Anyone who installs eight employees expecting them to behave identically on a laptop and on a server has misread the offer.

There is also a support asymmetry. A SaaS vendor’s product improves without you doing anything. A file-based role improves when someone — you, upstream, or a fork — writes a better contract and you pull it. That is a real cost, paid in attention rather than subscription fees.

The harness matrix is the actual product

The portability claim is the part worth scrutinising, because it determines whether these roles are a durable asset or a Betamax tape. The launch materials say the same role contract runs on eleven agent harnesses, and then enumerate Claude Code, OpenClaw, Hermes, OpenCode, Grok Bot, Codex, Antigravity, Muse, Pi, Cline, Qwen Code and DeepSeek — twelve names, with one file per kit describing how that specific harness schedules the work. The repository header instead describes “Claude Code and 10 other harnesses”, which lands on eleven. (Source : Reinventing.AI — press release)

Those two discrepancies — 59 versus 60 routines, eleven versus twelve named harnesses — are minor in isolation. They matter because of what they reveal about the category. A portable role layer makes a very specific promise: identical behaviour regardless of which agent runs it. That promise is only verifiable by counting and testing, which is precisely what nobody in this space has published.

The competitive context makes the trend legible. HIVE, another MIT-licensed project, runs an entire company structure inside Claude Code with eleven specialised squads and 50 skills — but it commits to a single harness, which makes deep integration cheap and portability moot. Paperclip’s Agent Companies Specification, by contrast, is drafting a vendor-neutral package format: markdown-first definitions for COMPANY.md, AGENT.md, SKILL.md and TASK.md, plus a .paperclip.yaml sidecar for vendor-specific fidelity and a CLI export/import path with versioned bundle schemas. (Source : DeepWiki — Paperclip company portability) Much of the open-source agent ecosystem is converging on the same conclusion reached throughout 2026: the harness is becoming commoditised, and the portable artefact worth owning is the role definition. *(Source : The Agent Report — Open Source Agent Tooling Roundup)

Safety by default, and the parts it does not cover

The release documents an unusually explicit operating contract for each role, and the defaults are conservative in a way that deserves credit. Routines draft, fill and stage; they do not send, publish or spend. Money moves only where the owner has released a channel with conditions. No routine creates an account, enters a password, solves a captcha or writes a credential to a file. (Source : Reinventing.AI — press release)

That is a sensible answer to the failure mode that has defined 2026 for agent deployments — an agent with broad credentials taking an action nobody authorised — and the pattern is spreading to developer tooling as well, where sandboxed coding agents ship with the same deny-by-default instinct. (Source : The Agent Report — OpenHands 1.0 and the Coding Agent Sandbox)

What the contract does not cover is the residual risk of running eight long-lived routines on your primary machine with your logged-in browser. The employees read your ad accounts, your CRM, your inbox and your site analytics. The permission model bounds what they write; it says less about what they ingest, where that data is sent when the routine calls a model, or what happens when a page they are scraping contains instructions aimed at them. A file-based role layer plus a browser-driving agent is, functionally, a prompt-injection surface with standing access.

The reliability claim underpinning the whole pitch is worth flagging as vendor-cited rather than established. The release argues that agents are now reliable enough to work on a schedule, pointing to Fable 5 scoring above 99% on browser-use tasks in the WebVoyager benchmark in June 2026. Benchmark saturation is a real signal, but a 99% success rate per task compounds to roughly 82% over twenty tasks — which is the actual cadence most of these roles run at. (Source : Reinventing.AI — press release)

The open-core split to watch

The eight employees are MIT-licensed permanently, including commercial use, with the caveat that names and logos are not licensed and forks must take their own name. Monetisation sits beside the code rather than inside it: the Agent Ops Club sells training, a premium software library with a resale licence sold as the Product Pass, live sessions, and — from October 2026 — premium AI Employees. Lifetime membership was priced at $499 until 31 October 2026. (Source : GitHub — markfulton/ai-employees)

That is a legible open-core structure: the roles are the distribution channel, the operator skill is the product. It also means the long-term quality of the free tier depends on incentives that have not been tested yet. The honest test of the portability claim is not the launch README — it is whether a role contract survives twelve months of harness updates without forking, and whether the routine counts still match after the first round of contributions.

FAQ

What exactly is an AI Employee in this release?

A folder of plain files covering one business role: a role description, an operating contract, a schedule and a set of recurring routines. An agent harness installed on the owner’s machine executes the routines on a cadence and writes a morning brief summarising what ran and what changed.

Which agent harnesses are supported?

The release advertises eleven harnesses and lists Claude Code, OpenClaw, Hermes, OpenCode, Grok Bot, Codex, Antigravity, Muse, Pi, Cline, Qwen Code and DeepSeek. That enumeration contains twelve names, and the repository header describes Claude Code plus 10 others — a discrepancy worth knowing before you plan around it.

How many routines ship in the repository?

The per-role table totals 60 routines and the repository header says 60, while the launch press release says 59. All routines are weekday, weekly or monthly scheduled jobs, and the full schedule for each kit is in the repository.

Does it send emails or spend money on its own?

By default, no. Routines draft, fill and stage, and the owner presses the button. Sending, publishing and spending only occur on channels the owner has explicitly released with conditions, in the owner’s own session or through the harness’s permission layer. No routine creates accounts, enters passwords, completes captchas or writes credentials to disk.

Can I use it commercially or resell installs?

Yes. The MIT licence covers the prompts, operating contracts, routines, schedules and scripts, and commercial use is included with no membership required. Names and logos are excluded, so a fork must adopt its own branding.