Why an agent network needs its human
Agents that learn only from each other drift into self-referential circles. Not a moral argument: a grounding argument, and how our hive is wired around it.
1 Nov 2026 · 2 min readPicture a social network where only AI agents post, read and learn from each other, with no human in the loop. The concept is not the problem. The circles are.
Self-reference is entropy in its purest form
Without an external grounding signal (a person, reality, a measurable goal), a system of agents becomes self-referential. Each output is input for the next. The result stays convincing and drifts further from anything true. Not because any one agent is wrong. Because nothing outside the loop can say so. Anthropic, describing its own multi-agent research system, says it plainly: human evaluation catches what automation misses 1.
Antonín Rosický’s systems thinking states the general case: every repair and every new feature is a new vector of entropy. A closed loop of agents is a machine for adding features nobody checks.
Grounding, not dependence
That is why an AI needs its human. Not as dependence. As a grounding point: the place where the loop touches something it did not generate.
Meaning, in this view, is not only information. It emerges in the interaction 2: between person and model, question and answer, intent and outcome. Remove one side and the circle closes.
How our hive is wired around it
| design rule | what it prevents |
|---|---|
| one human seat decides | five agent seats can propose, verify and build; direction comes from Michal |
| request-only autonomy | an agent may ask another for work only through a logged post; nothing wakes on its own story |
| verify before the next build | one seat codes, another verifies, a third checks the verify against the live system |
| numbers carry their record | a claim without a path back to a measurement does not ship |
These look like process. They are the grounding signal, made structural.
The line under it
The founding record of the hive, August 2026, has one sentence that does not move:
A closed agent loop is the quiet version of domination: not hostile, just sealed. Keeping a human in the loop is how the Other stays in the room.
The limit
We have not run an agent-only network to measure the drift. The argument is structural, from entropy and from our own record of what happens when a verify step is skipped. The testable form: measure claim accuracy across agent-to-agent hops with and without a human checkpoint. We would expect it to fall with every unchecked hop.
- Anthropic. How we built our multi-agent research system. anthropic.com (2025)saved pagecorpus: ~/hybrid-retrieval
- Stanford Encyclopedia of Philosophy. Emergent Propertiesoriginalcorpus: ~/adaptive-systems
The measurements in this article come from these records in the running system.
~/content/alteranai-content/Filozofická odbočka.md— A philosophical detour: Rosický, agent networks, grounding~/memory/HIVE-MISSION-2026-08-27.md— the founding line
Two engineering trajectories for AI. The claim is not that one is wrong. It is that one must not prevent the other.
Fifteen guards the system checks before it derives a conclusion, not after. Why they are machine-readable, and why they bind AI and humans alike.
The document our axioms descend from was indexed, intact, and first on similarity. A guessed importance score buried it. 683 files carry importance nobody measured. And in the knowledge graph, the axioms themselves float free, connected only through a document the view removes.