Multiplayer AI Should Not Require Your Team Using GitHub
Emily Kramer recently published a smart framework for getting marketing teams out of “single-player Claude mode.” Her main point is right. The next big jump in AI productivity won’t come from one keen marketer getting even better at prompting. It will come from taking what that person has learned and turning it into something the whole team shares.
She’s right about where teams should end up, and she’s honest about what it costs to get there. Our view is that the cost is the problem.
The destination: compounding momentum
Right now, most useful AI work stays with the person who did it. One marketer works out a good research process. Another builds a content workflow that reliably works. Someone else finally gives Claude enough background to write an accurate sales brief. Each of them gets faster. The team doesn’t necessarily get any better.
Multiplayer AI fixes that. Context is shared, the useful capabilities spread, and a correction one person makes improves the next answer for everyone. The team stops starting from zero and builds on its own past work, so the gains add up over time.
Is there evidence it’s worth doing?
We recently analyzed 3.8 million messages on elvex to see what happens as a multiplayer harness takes over more of the assembly and execution. The median message got 23% shorter, from 164 to 127 characters. Tasks got harder, with requests per conversation up 60%, from 5.25 to 8.42. Satisfaction rose from roughly 50% to 79%.
The users didn’t turn into expert prompt engineers overnight. The harness was doing more of the work for them.
That’s what a multiplayer system should produce: people giving fewer step-by-step instructions while getting harder work done. The knowledge needed for a good result moves out of each person’s prompt and into infrastructure the company shares.
The cost of getting there
Kramer has hold of something important, and she also shows why so few teams have gotten there: building a shared AI system takes a lot of work.
Her recommended setup uses Claude, Claude Code, Cowork, MCPs, a shared GitHub repository, plugins, skills, context files, publishing processes, maintenance routines and a named owner. Each of those is reasonable on its own. GitHub isn’t even that hard once someone walks you through it.
But you are not going to get a whole marketing team using GitHub. A few people will. Most won’t. They have campaigns to launch, copy to review, reports to write and pipeline targets to hit. Nobody joined marketing to manage local repositories, sort out versioning questions or decide whether a fix from a chat should be committed to a shared skill file.
Now ask the same of sales. Then accounting. Then HR, legal, operations, customer support and every other nontechnical team.
If multiplayer AI only works when the whole company learns developer workflows, it will stay with small groups of very motivated enthusiasts.
The goal is right. The path to it asks too much of the people it’s supposed to help. Luckily, companies don’t have to build multiplayer AI that way. The collaboration layer can be a product.
From individual productivity to organizational capability
The first wave of AI at work was nearly all single-player. People opened ChatGPT or Claude, typed a request and got an answer. The best users learned to add more instructions, upload better documents and write elaborate prompts. They built personal workflows that saved them hours.
A senior marketer could spend months teaching Claude how the company talks, what the product does, which customers matter and how a good campaign brief is laid out. A colleague who opened a new conversation still started with none of it. The senior marketer got better at using AI.
The company, though, never got to keep that marketer’s method. That’s the difference between adoption and transformation.
Adoption means lots of employees have access to AI. Transformation means the organization has a new and better way of working. The surest route there is an AI habit that improves every time people use it, and a multiplayer system is what makes that possible.
Then the best person on the team does more than turn out better work. Their expertise raises the starting point for everyone else.
That’s the opportunity Kramer points to, and it goes well past marketing. Sales, operations, HR and finance can all gain from it.
The DIY version reveals what the product needs to do
Kramer lists four ingredients for multiplayer Claude: Claude, context, capabilities and collaboration. Her list is a good way to see why a shared account or a folder of prompts falls short.
A real multiplayer system needs up-to-date company knowledge. It needs repeatable ways to do the work and a way to hand improvements out to everyone. It needs owners, testing, upkeep and trust. In a DIY setup, the team has to put all of that together itself.
Someone has to decide where the official context lives. Someone has to connect the right tools and data. Someone has to turn work that went well into reusable skills, and those skills then need testing, publishing, updating and auditing. Everyone’s local setup has to be compatible. Changes have to reach everyone. Duplicate and outdated capabilities have to be found. Permissions and sensitive information have to be managed on top of all that.
For a team that’s comfortable with technical work, that’s a fair experiment. As the way a whole company runs, it doesn’t hold up.
It also creates a familiar organizational problem. The person who knows how to build the system usually isn’t the person running the department. The department head understands the work but maybe not the infrastructure. Everyone agrees a shared system would help, and nobody has time to stop working long enough to build one and keep it running.
The harness is the product: here’s how we think about harnesses
A model supplies intelligence. A harness turns that intelligence into useful work.
Ask a bare model to “write a sales follow-up email” and you’ll get something generic. Give the same model the CRM record, the meeting transcript, the relevant product information, approved messaging, earlier emails and a clear way to review and send the result, and the same short request produces an email you can use right away.
What changed is everything the harness put around the prompt. The user’s prompt stayed the same.
For each request, a good harness works out what the model needs to know, which tools it can use, what the person is allowed to see, what’s already been done and what should happen next. It finds the relevant information so the employee doesn’t have to rebuild the company’s knowledge in every conversation.
We find it useful to think about this system as five connected ideas:
Models are commodities that provide intelligence
The model does the reasoning. It reads the request, works through the task, and writes or chooses the next step.
Models matter, but they’re one input to the system. Different jobs need different amounts and kinds of intelligence, and the best choice will keep changing as new models come out. A company’s workflows shouldn’t have to change every time.
Above all, companies shouldn’t pay full price for frontier models like Claude and ChatGPT when other models do the job just as well for 5% of the cost.
The harness automatically collects context
An employee shouldn’t have to tell the model everything the company already knows.
A harness can gather and keep context from the person’s role, their current work, past conversations, shared team knowledge, company systems, connected tools and the resources they can use. Then it pulls out what’s relevant to the task in front of it.
That doesn’t mean cramming every document into every prompt. More context isn’t automatically better. Context that doesn’t apply costs money and distracts the model.
The harness should keep information on hand, understand how it relates to the person and the task, and bring it in when it helps. The employee types less because the harness does more of the assembly.
Resources are singular pieces of context for a purpose
Resources are the individual parts the harness can draw on.
An agent can hold the instructions and behavior for a particular job. A datasource can supply company knowledge. An integration can fetch information from another system or take action in it. An artifact can keep work the company has already done. Other resources can hold settings, tools, evaluations or material for a specific task.
Each resource has a defined purpose. You make it once, improve it over time and offer it wherever it’s relevant. And it lives in the cloud, so when someone saves a change, everyone gets the update automatically.
This is the product version of the skill files, context documents, connectors and workflows that teams otherwise manage by hand.
Shared spaces bring resources and people together
A useful AI system has to understand that people work with colleagues toward shared goals.
A shared space is where resources come together for a team, project or department. A marketing space might hold the company’s messaging, customer research, content agents, the CRM integration, campaign artifacts and current plans. The people who own that work get access to the space and what’s in it.
A sales space would hold a different mix. So would HR, accounting, legal or a cross-functional team launching a product.
That makes context multiplayer without making all of it available to everyone. People get the resources their work needs, within limits that fit their role.
When someone adds a useful resource to a space or improves one that’s there, everyone else with access benefits. The shared system gets better without anyone having to pull a repository, install a plugin or update a local configuration.
The agent can pick the right resource for each task from inside that managed Space.
In practice, your employees don’t have to remember where 30 different agents, datasources and other resources live. You point them to their team’s Space, where it’s already organized for them. That takes a real load off their minds.
Control exists at every layer
Multiplayer without control turns into chaos.
Companies need control at the level of the company, the person, the resource and the space. They need to see what people and agents are doing. They need permissions that set who can see, use, edit or share each resource. They need governance over connected data and actions, and a record of how the system is used and where it’s paying off.
They also need a say in what they spend.
Control shouldn’t mean forcing everyone into a rigid workflow designed at headquarters. It should set safe limits inside which people can experiment, build and share what works. The best systems make the approved path easier than shadow AI, and not only stricter.
A harness can do this where a pile of local files and personal subscriptions can’t. It gives you one environment where experimenting and oversight support each other.
Model independence makes multiplayer affordable
Building the shared system around one model provider has another problem: every workflow that succeeds gets tied to that provider’s prices.
For the last few years, defaulting to the newest frontier model made sense. The quality gaps were real, nobody knew which uses would matter, and the simplest advice was to use the best model available for everything.
Today that habit wastes a huge amount of money.
Summarizing a meeting doesn’t always need a frontier model. Neither does sorting a support request, drafting an internal FAQ, looking up a policy, reworking a document or following a clearly defined workflow with good context.
Open-weight models can now handle much of the everyday work people give AI. With an independent harness, companies can get that work done and spend 95% less on tokens.
That’s more than a purchasing win. It changes what a company can try.
When every experiment runs on the most expensive model there is, teams get cautious. Managers restrict access, promising workflows stay stuck in small pilots, employees worry about usage, and the workflows that succeed cause budget worries just when the company should be expanding them.
At 95% lower token cost, those limits move.
Teams can run more experiments and more people can take part. Agents can take on longer, multi-step work. A workflow that was too expensive to run on every customer account can start to make sense. Something one team has proven can spread to ten more without blowing the budget.
Lower costs give the company room to find out where AI really pays off. Most experiments won’t change the business, which is exactly why they need to be cheap. The company can try more ideas, learn faster, drop the weak workflows and expand the ones that work.
That gives you a healthier adoption cycle:
- More people can experiment.
- Work that succeeds becomes a shared resource.
- Shared spaces get that resource to the right groups.
- Use produces more corrections and more company context.
- The harness folds those back into the system.
- Lower model costs pay for the next round of experiments.
The value and the momentum keep building, and the cost of each extra use stays manageable.
The harness can make a less expensive model more useful
Model comparisons often treat the prompt as the whole input. Give two models the same instruction in isolation, compare what they write and declare a winner. Useful AI at work doesn’t run like that.
A model inside a good harness gets the relevant company context, the right tools, examples of work that was accepted, knowledge of the user’s role and a workflow built for the task. It doesn’t have to guess the whole situation from a few sentences.
A pricier model without that context can do worse work than a cheaper model inside a well-built harness. That’s why, for many enterprise tasks, the model is turning into a commodity input. It still matters, but much of the practical value comes from the system that points it at the job.
The good news is that this already exists: try elvex
Emily Kramer is right that single-player AI isn’t enough, and that the best person on a team should make everyone else better.
The next step is to build multiplayer into the product itself.
The model is the commodity input. The harness is where the company’s context, capabilities, collaboration and control stick around for good, and that’s the point at which AI starts producing results across the whole company.