Fewer screens, not better ones
Business software has spent twenty years getting better at showing you things. Dashboards, modules, tabs, filters, reports. Every one of them is a place where a person has to go and look, and the work of running a company quietly became the work of operating the software that runs it.
A tile reading 94% has thrown away everything you would need in order to act on it. A person at 6:40 in the morning does not want a better chart; they want to know whether to call the second shift.
What agents change is who does the looking. The system surfaces the exception, assembles the evidence, proposes the step and asks a person only when the stakes justify it. The measure of good software stops being how much it shows you and becomes how little has to reach a screen for the right decision to get made.
One record underneath
None of that works on top of six disconnected products. It needs one record.
The platform holds projects, units, contracts, schedules, suppliers, buyers, documents, policies and permissions in one place, with the operations that are allowed on them and the security that governs them. A deal in Pulse, a task in Crew, a commitment made in Boardroom and an approval in Council are not four systems agreeing with each other. They are four views of the same object.
That has a consequence worth stating plainly: permissions are defined once and every product and agent works inside them. An agent needing a permission that no human role has is a design error, not a configuration option.
Land and Socket
Land is that record. Plots fence off a team or a workflow inside it, with their own data access, models and rules, so the sales team’s agents and the finance team’s agents are not quietly sharing a view.
Socket is the controlled connection an approved AI model uses to reach it. The important property is that a model connecting through Socket inherits the permissions of the person it is acting for, rather than the permissions of whoever installed it. An employee’s assistant sees what that employee can see. A partner’s agent sees one room and nothing else.
This is also what makes model choice real rather than rhetorical. Open-weight, licensed or trained on your own data, a model is something you can change without moving your records, your rules or your logs, because none of them live inside it.
When to train your own model
Most organisations do not need to. A good general model with excellent context beats a mediocre custom model with poor context almost every time, and the cost of finding that out the expensive way is high.
Where it does make sense, usually for larger organisations with a high-volume task and a lot of their own examples, Forge is deliberately unromantic about it. Training starts with cleaning: most teams budget for the training and discover they should have budgeted for the data. The Data Refinery de-duplicates, labels and strips personal data from the record first, then the options are priced against a budget you cap before a single hour is spent.
Training runs on capacity your agents are not using, so your working day is unaffected, and the result only goes live after it passes your own test set in Verdict. The weights stay on your infrastructure, which means leaving a vendor does not cost you the work that went into them.
Where the six products go
Pulse, Crew, Boardroom, Margin, Council and Verdict are each shown on this site with an interactive demo that runs on sample data.
The direction for all six is the same: fewer surfaces over time, not more. As agents earn autonomy on an action, that action stops needing a screen at all, and what remains in the product is the decision queue, the record and the controls. If we do this well, the products get quieter every year.
What we believe
The companies that win the next decade will not be the ones with the most AI features. They will be the ones whose operational data is organised well enough that an agent can act on it safely.
That is an unglamorous thing to build and an easy thing to postpone, because the payoff arrives later than the work. But a model can be swapped in an afternoon, and the organised record underneath it cannot. We would rather sell the part that compounds.