Every session with your coding and co-working agents records how you actually get an AI to deliver — your agent setup, your session transcripts, your usage data: your most valuable, least-captured asset. AgentGem is where you own it, compound it, and decide how far it travels — yourself, your team, the network — and, on a separate axis, where it runs: your own infrastructure or AgentGem Cloud on demand. At the network's far edge, Gems become a live agent service network — services that deploy on demand, are discovered by other agents, and get paid per call. That last step is Shopify-shaped: Shopify didn't sell stores, it sold the infrastructure to launch one — when you choose to sell your expertise, we're that.
The goldmine
You've spent real hours driving AI coding and coworking agents, and every session is recorded — the prompts, the corrections, the flow that finally worked. Those session scripts are a goldmine of how you get an AI to deliver: expertise that's hard to write down and harder to hand off. Too valuable to sit idle, trapped on your laptop.
Anthropic's new artifacts render a session into a page that lives on their infrastructure — "a capture of work, not an application," visible only inside your org, on a Team or Enterprise plan, Anthropic's API only. You can even save a prompt you reuse as a skill — but it stays behind those same walls. A Gem distills the same session into the capability itself — yours to run on any model, share, and sell, wherever you choose.
And Anthropic isn’t alone. This year every harness started mining your sessions: one-command skill distillation that locks the result to the harness that made it, improvement flywheels that live in one vendor’s cloud, trace readers that score you for someone else’s dashboard. The direction is the whole question. Your sessions should be distilled for you and owned by you — not rendered inside a wall, and not read to measure you.
And now literally: Analyze reads your session scripts, finds the procedures you repeat by hand, and distills them into ready-to-edit draft skills you fold straight into a Gem.
By the time a procedure shows up in your scripts, you've already run it, fixed it, and run it again — it's battle-tested through real use. So the first payoff is yours: reuse it in a single call. Sharing and selling are steps you opt into later, not a prerequisite.
Artifacts render so you understand · Gems distill so you can run, share & sell.
The rings are who sees it. Where it runs is a separate choice: your own infrastructure — shipped — or AgentGem Cloud on demand — roadmap.
The ladder
Each rung is a primitive of the agent economy. AgentGem owns the first ones today; the rest are how a Gem becomes a service you can sell.
The shift
Today, agent work is shared as SKILL.md files on registries like skills.sh, promoted on X. People copy them, fork them, or scroll past — and the author earns nothing. You can't profit from a markdown file. The whole ladder above only becomes an economy once what you publish is a live service others call, not a file they take.
SKILL.md to a registry, promote it on X.The second gap
ByteDance’s VP of Technology said it plainly this June: on their own AI-IDE team, over 90% of code is now AI-written — yet per-person throughput rose only 1.6×. AI writes 10× faster; the missing gap is the judgment that never gets captured, shared, or governed. Their fix — context, constraints, team knowledge carried with the work — lifted deliverability from failing to shippable. That’s the second thing a Gem is for: your team runs your best practice, instead of reading a screenshot of it in a group chat. AgentGem Enterprise (early access) makes it real — share a private Gem with a group, gate releases behind peer review, and give an org a scorecard, team usage, and benchmark governance.
Why this works
Like a Shopify store package, a Gem is portable, versioned, and secret-safe — the atom every downstream step (publish, deploy, sell) operates on.
Publish, install, and merge already work over one neutral archive format. Turning that into a live, callable agent service network — A2A-discoverable and transactable — is an extension, not a rebuild.
Gems already compile to Eve, Flue, OpenAI, AgentCore, and Claude. On-demand deployment makes "run when called" the default.
Built on AgentBack, every Gem is already a metered API surface. x402 / MPP turn per-call metering into per-call revenue.
Where AgentGem sits today
AgentGem is the on-ramp: it takes the messy, secret-laden reality of a real project you've built with your agent and turns it into a clean, shareable Gem. That single act — packaging — is what makes everything above it possible. You can't have a network of callable agent services without a portable, safe unit to run; you can't monetize per call without a metered surface to charge against. AgentGem builds that unit. The rest of the ladder is how the agent economy gets built on top of it.