A tool executes. An employee decides.

Nobody calls a calculator a colleague, however brilliant it is. You give a tool an instruction and you check what comes back. You give an employee a goal, and you get back decisions you never had to make yourself. What you pay an employee for is judgment you can rely on.

Today’s agents are the most capable tools ever built. They write, code, research and operate software at a speed no person can match. Now ask them what you would ask any new hire. Will you make the call we would make? Do you know when to stop and ask? Can you show me why? Will you be better next quarter because of what happened this quarter?

While the answer is no, every agent needs a person standing behind it, and that person is the bottleneck. The agent stays a tool.

Software is being rebuilt for the third time.

For forty years, software recorded what happened, then moved work between people. The judgment stayed in someone’s head. The third generation has to hold the judgment itself: why a decision was made, on what evidence, what it produced, and how to decide better next time.

Diagram 6 · Five layers, one closed loop. From the bottom: Record, Context, Decision, Action, Review and learn. The Decision layer is the core. Outcomes write back, so the next decision is better.

We believe every company will soon run more agents than people. The winners will not be the ones with the most agents or the smartest models. They will be the ones whose agents decide the way the company would, and decide better every week. Decision intelligence is what turns a fleet of tools into a workforce.

Models get cheaper and stronger every quarter, and software itself is getting cheaper to build. What stays scarce is knowing what a good decision looks like here: in this industry, at this company, on this kind of call.

01 · A decision is not a clean chain.

On a slide, a decision is three boxes: context, decision, outcome. Inside a company it never is. Say a team stops Campaign A. The reason on record is that its return fell below the benchmark. The real reasons also include a cash target for the month, a second campaign with better marginal return, a client who wants no surprises at quarter end, tracking that has started to fail, and a creative team with no time for another test. Some of this sits in a database. Some was said in a meeting. Some was never said at all, because everyone in the room already knew.

A system that records only “return was too low” turns a portfolio decision into a rule, and the rule is wrong. And when the second campaign improves afterwards, was it the budget, the new creative, the season, or a competitor pulling back? An honest system keeps the stated reason, the evidence, the outcome and the proven cause as four separate things, and it is allowed to say “we don’t know yet”.

Diagram 7 · A decision is not a clean chain. On a slide, a decision is three boxes. In a real company, many signals feed several decisions, some of them never written down; decisions and outside events produce outcomes together; and every outcome becomes the next context.

02 · The same engine means something different in every vertical.

The plumbing can be shared. The meaning cannot. Here is the same decision system in two verticals.

Different objects are only the surface. Underneath, the standard of evidence, the limit of risk and the definition of success all change. Even inside one vertical, stopping a campaign, approving a budget and investigating attribution each need their own agreement on what must be known, which source wins and who may act.

03 · A general agent can learn what you did. It cannot learn what you should do.

Give a general agent every file, log and message a company owns, and it will infer how that company has decided. That is a real threat to thin software, and we take it seriously. But a record of the past is not a standard. The agent cannot tell a policy from one person’s exception. It cannot settle which system holds the truth, grant itself the right to act, or answer for the result. Those are earned inside real work, with real outcomes, over time.

04 · So we build it one decision at a time.

Vertical, then use case, then one decision. We start with a decision that is frequent, measurable and expensive, such as “should we stop this campaign?”, and close the whole loop around it: the evidence, the call, the action, the outcome, the lesson. When that loop holds, the same engine takes the next decision, and later the next vertical. Our first vertical is growth, through MegaHit.

This is slow on purpose. Ten thousand decisions with their outcomes teach more than a million documents without them, and nobody can buy those decisions in a labeling project. They have to be lived.

Diagram 8 · One engine, many verticals. A shared engine at the centre. Growth is the first vertical and is live. Each further vertical is a new orbit whose judgment has to be earned.

Ten principles we build on.

1. Start from the decision, not the model. Pick one frequent, measurable, expensive decision and build outward from it.
2. Memory is not a chat log. Useful memory is selected, structured, versioned, permissioned and forgettable. Dumping text into an index is not memory.
3. A reason is not a cause. We store the stated reason, the evidence and the outcome separately, and never promote a reason to a proven cause.
4. The closed loop beats data volume. Data only teaches when it is tied to a judgment and a result.
5. Retrieval is itself judgment. Which precedent applies depends on the situation, not on matching keywords. Choosing it well is industry knowledge.
6. No outcome, no automation. If you cannot measure whether a decision worked, you are not ready to automate it.
7. Every action needs a leash. Automated actions carry permissions, an audit trail and an undo. Authority is granted in tiers, never all at once.
8. Horizontal platform, vertical semantics. The engine is shared. The ontology, the thresholds and the judgment are specific to each industry and each company.
9. Organizations learn, not just agents. One agent’s win should become every agent’s default by the next morning, without breaking what already works.
10. The goal is better judgment, not better chat. The interface may be a chat, an approval, or a workflow that runs on its own. What matters is that the organization decides better.

Super Intelligence = Human Intelligence + Artificial Intelligence.

The industry is racing to build more AI: more agents, more memory, more monitoring. We think the term missing from the equation is the human one.

Humans set the standard. AI carries it everywhere. Each makes the other better. We are not waiting for a model that no longer needs people. We are building the system in which people and models become more than either one alone.

Each player manages a slice. We manage the whole team.

The 2026 market answered “who runs our agents?” in six different ways. Each is useful, and each covers one part of the job. As we read them, most add more artificial intelligence to manage artificial intelligence. Super Intelligence adds the human term and makes it scale.

Diagram 10 · Positioning map. Columns show what is managed, from access to judgment. Rows show who learns. Super Intelligence is alone where judgment meets organizational learning.

The seven jobs of a manager, company by company.

Company judgment · Context sync · Performance review · Strategy adjustment · Shared experience · Token visibility · Report up.

Diagram 11 · The seven jobs of a manager. Based on each company’s public materials as of October 2026. The Super Intelligence row shows our design.

How we differ, and when they fit better.

The most valuable software is not a database of company facts. It is infrastructure that compounds a company’s judgment.

— Kisson Lin · CEO, Super Intelligence