QpiAI Agent HiveWorld models for operations.
World models for operations.
Agents that reason and act inside them.
We model how an organisation runs, from its records and from the people who know them. Then we build agents that make decisions inside that model and cite every step.
Why did first-pass yield dip on line 3 last week?
- Readschema · production.batches0.4s
- Queried4 tables · 12,480 rows1.9s
- Comparedweek 41 against week 400.6s
- Checkedmaintenance_log, shift_roster0.8s
- Wroteanswer with 3 sources0.3s
Answer3 sources
−3.6 pts87.6%
First-pass yield, line 3, week 41
Most of the drop is on the night shift after Tuesday's fixture change. Lines 1, 2 and 4 held steady.
batchesmaintenance_log #2231shift_roster
Ask a follow-up…↵
Platform
What the product does. Context, workflows, review and evals in one place.
Context from where work lives. Connect a system once. Every agent reads the same picture.
ConnectorsWorkflows you can rerun. Steps, checks and hand-offs that run every week.
WorkflowsThe ontology changes by proposal. A definition moves only when its owner approves.
ReviewBetter with each release. A version ships only when it beats the last one.
Evals
Figures on this page are illustrative.
Context factory · long-running agents
One graph of how you work. Agents that keep working on it for days.
Your whole catalog, as one graph. Modules, schemas, tables and KPIs, with the lineage between them. Watch the tour, or open it and explore.
How the context factory worksRuns that last for days. Agents checkpoint, wait for approvals and resume where they stopped.
How runs workWhat we work on
Seven research directions. Each one is in the product today.
- Research directions01 / 09
How Agent Hive learns your operation. Seven ideas underneath.
Agent Hive - Research directions02 / 09Step 01
A context factory
Sources become reviewed, typed pages in one graph.
- Research directions03 / 09Step 02
An ontology with room
Hierarchies live naturally in hyperbolic space.
- Research directions04 / 09Step 03
Swarms on the graph
Many small agents; the useful paths are reinforced.
- Research directions05 / 09Step 04
Runs that last days
Checkpoints, resumes and hand-offs.
- Research directions06 / 09Step 05
Self-improvement
Edits to the agent configuration, graded and approved.
- Research directions07 / 09Step 06
Small models, fine-tuned
Task-specialised SLMs that run on your own hardware.
- Research directions08 / 09Step 07
Verified before it runs
Proofs, or the shortest counterexample.
- Research directions09 / 09
See it on your data.
Contact usAgent Hive
Latest research
Notes from the work. Written by the people doing it.
- GitOfThoughts: version-controlled reasoningWe saved an agent’s reasoning tree in an ordinary git repository, one commit per scored thought. Memory helped only when a stored trace was nearly a copy of the new problem, and git surfaced every contradictory lesson as a merge conflict.
- Retrying isn’t the same as learning why you failedSampling again or asking a model to reflect tends to repeat the mistake. ReTreVal labels why a branch failed and retries from a different strategy: on MMLU-Pro, with no answer key in the loop, that added 10.8 points.
Industries
Applied where the rules are strict. From plants and fabs to banks and networks.
ManufacturingQuality deviations traced from the control chart to the corrective action.AutomotiveField issues correlated across signals, software, suppliers and service.SemiconductorYield excursions traced through lots, tools and metrology to a disposition.EngineeringDrawings, certificates, test and simulation evidence behind each decision.BFSIRegulatory change, withholding tax and collateral, checked against the rules.FMCGEvery order exception explained across allocation, stock and delivery.TelecomNetwork changes checked against live state, vendor rules and service impact.HealthcareOperational quality measures and policy, investigated with sources.
Bring us one hard problem. We reply within one working day.
agenthive@qpiai.tech