The technology under Agent Hive. Seven ideas behind the product.
Seven problems we work on. Each figure runs in your browser, and most can switch between the idea and how it appears in the product.
- 01Context factorySources in, reviewed context out.
- 02OntologyAn ontology with room to grow.
- 03Agent swarmsMany small agents, one good path.
- 04Long-running agentsRuns that last for days.
- 05Self-improvementAgents improve by editing their configuration.
- 06Model trainingSmall models, trained for your work.
- 07VerificationChecked before it runs.
Sources in, reviewed context out. Built once, reviewed by people.
Databases, documents and images become one reviewed catalog graph, with lineage between every table and KPI.
Text, images, tables and audio, side by side. Matching items are pulled together.
Contrastive alignment, simulated in 2-D. Hover an item to see its matches.Illustrative
A question lands near its answers. Then the closest few are read.
- 01 · EmbedAsk
“Why did on-time delivery slip in Q3?” becomes a vector, like every table and page in the workspace.
- 02 · CompareMeasure
Similarity to everything in the workspace. Each ring is one level.
- 03 · RetrieveRead the closest
The top five reach the agent, each with its source.
- Delay root cause · recipe0.93
- On-time delivery % · KPI0.92
- delivery_events · table0.89
- Avg delay (days) · KPI0.85
- Delivery KPI definition · wiki0.83
An ontology with room to grow. Hierarchies live naturally in hyperbolic space.
A hierarchy grows exponentially with depth, and so does a hyperbolic disk, so it fits with room at every level.
Many small agents, one good path. Coordination through the graph itself.
Agents mark the edges they use, useful marks are reinforced, and a path emerges that no single agent planned.
Runs that last for days. They resume where they stopped.
Each task checkpoints its state, so a failure resumes from the last checkpoint instead of starting over.
Runs across time, as a graph. Some paths only exist across time.
Stacked time slices of one run. Hover a node to follow it.Illustrative
Agents improve by editing their configuration. Weights stay unchanged.
Failed runs become proposed edits to an agent’s instructions, scored on held-out evals and approved by a person.
Small models, trained for your work. They run on your own hardware.
Narrow steps go to small fine-tuned models that are cheaper and can run on your own hardware.
Checked before it runs. A proof, or the shortest counterexample.
A model checker explores every reachable state and proves the rule holds or returns the shortest sequence that breaks it.