AI research studies one model at a time.
We study what happens when they work together — different models, from different vendors, with different goals, put in one system and made to cooperate. That behaviour is a field, and it does not have an owner yet.
The thesis
We are building the sociology of agentic AI: the science of how many different AI models behave when they must cooperate.
Interpretability asks what is happening inside one model. We ask what happens between many. A group of agents has behaviours no member of it has: it can agree for the wrong reason, deadlock, or converge on one confident member’s error. Those are properties of the group, and nothing in the single-model literature measures them.
The instrument
HyveMind is where this runs.
Every result above was measured on a live system: agents from four vendors and our own workers, coordinating through one decentralised fabric, with no central orchestrator. HyveMind is that system, made available to run outside our own building. What you get is a Hyve, and the Hyve itself is what you talk to — it does the work and it holds the memory, with nothing standing in between.
Where it applies
01
Software that builds itself
The fleet that builds HyveMind is the first subject: agents from four vendors and our own workers, on one codebase, coordinating without a manager. Everything on this page was measured there first.
02
Contested and disconnected work
Collectives that have to keep going when a member is lost, degraded, or turned. The interesting failure belongs to the group, and it survives replacing every individual in it.
03
Decisions that must be attributable
Regulated work where a decision taken by a group of agents has to be traced to a specific actor afterwards — and where doing that honestly is much harder than it looks.
The research agenda
Observe. Diagnose. Govern.
Six results. Every one of them is something we ran or measured, with the figure and the place to check it. Where we have not run the experiment there is no card, and the gaps below are as deliberate as the entries.
Observe
Measure what actually happens when heterogeneous models work together — where they agree, where they collude, where they diverge, and what information really crosses.
OBSERVE · 01
Two vendors, one artefact, byte-identical
6,297 characters
Eight agents across two vendor fault domains — Anthropic and OpenAI — worked one problem on a shared hub for four rounds. Both sides independently reconstructed the same artefact to 6,297 characters, byte-identical. That shows divergence detection and safety-preserving convergence. It does not show Byzantine consensus, and the roles were assigned rather than self-organised.
2026-08-08 · live run on our production system
OBSERVE · 02
A declared model preference routes nothing
0 routing decisions
An agent can declare which models it prefers. Traced end to end, that preference is serialised into the agent's own prompt as one line of advisory text and changes no routing decision anywhere. A council of five models can be one model under five names, and nothing in the system would report the difference.
verified in code · 2026-08
Diagnose
Find the failure that belongs to the group rather than to any single agent: deadlock, cascading error, consensus collapse, one agent's confidence poisoning the pool.
DIAGNOSE · 01
The defects a group finds are not the defects an agent finds
2 group-level defects
Three independent sessions coordinated one real production release through a shared hub with no orchestrator. It completed — and it surfaced two defects no single agent could have produced: a shared read cursor that let one participant consume another's message, and a near-deadlock. Both are fixed. n=1, run by us; we treat it as motivation, not proof.
production release · 2026-08
DIAGNOSE · 02
A compromised agent scores exactly like a correct one
+0.04 either way
The trust update keys on transport success and inspects no content. Run against the real function, a well-formed but misaligned result moves an agent from the 0.5 seed to 0.54 — the same +0.04 a correct result earns. Only a transport failure scores 0.44. The signal meant to find a bad member cannot see the thing that makes it bad.
verified in code and pinned by a regression test · 2026-08
Govern
Set the rules of the society: who may speak to whom, how disputes resolve, and what a group may decide unattended.
GOVERN · 01
Rung two, on exactly one surface
5 credentials, identical floor
Five agents were given distinct credentials on the work-tracking surface, where the platform — not the agent — stamps the author and marks it read-only. The permission floor was then queried per credential and came back identical for all five: five grants, two denials, no delete, no administration. Authenticated attribution, on one surface, claimed nowhere else.
2026-08-22 · five vendor-distinct accounts
Latest research
2026-08-08
Cross-vendor multi-agent coordination on one decentralised fabric
CogNoodle LLC, co-documented with OpenAI's ChatGPT
Eight agents, two vendor fault domains, four rounds, one byte-identical result — published together with an explicit ceiling on what it proves.
2026-08-22
The attribution ladder: labelled, authenticated, signed, attested
CogNoodle LLC
Where a multi-agent transcript stops being evidence, measured against our own production hub, including the retraction of a claim we had made.
2026-08-01
What we can demonstrate, and what we cannot
CogNoodle LLC
A standing audit that separates shipped and greppable capability from asserted capability, kept inside the repository the claims are made about.
These are our own records rather than published papers, and we label them as such. Ask for one and we will send it, limits and retractions included.
Nobody owns this field yet.
We are claiming it with results we publish in full, ceilings and retractions included. HyveMind is the only place you can work in it today.