Interloom.
See what happens next.
Specialized AI agents reason across interconnected systems and expose the cascading consequences of a single event or policy decision — turned into an inspectable graph rather than an opaque answer.
Can specialized AI agents reveal what one model or one analyst cannot hold at once?
Every consequential decision — a drought forecast, a policy shift, a supply shock — travels through interconnected systems that no single expert can hold in full view. Interloom asks whether a chorus of bounded, specialized agents can make those cross-domain consequences legible, inspectable, and open to challenge.
Structured handoffs beat undifferentiated answers.
Structured causal handoffs between bounded domain agents will produce more legible and inspectable system-level reasoning than an undifferentiated assistant response. Each agent commits to a typed claim; each claim can be inspected, disputed, or discarded by a human reviewer.
The interface should expose the seams between agents, not hide them behind a single confident voice.
Scenario intake → domain agents → typed consequence claims → critique → cascade graph → synthesis.
The proposed architecture routes a scenario through six named stages. Evidence, provenance, uncertainty, dissent, audit, and human review cut across every layer — no stage operates without them.
- Stage 01Scenario intakeA single event, decision, or policy is framed as a structured input with scope and constraints.
- Stage 02Domain agentsBounded specialists reason inside their own remit — water, agriculture, biodiversity, and so on.
- Stage 03Typed consequence claimsEach agent emits claims in a shared schema: statement, evidence, uncertainty, and dependencies.
- Stage 04Critique & reconciliationPeer agents challenge, refine, or reject claims; dissent is recorded, never silently resolved.
- Stage 05Cascade graphSurviving claims compose into an inspectable directed graph of cross-domain consequences.
- Stage 06Decision synthesisA human reviewer reads the graph, weighs uncertainty, and produces the actionable synthesis.
Every claim points back to the sources or reasoning that produced it.
Which agent authored what, when, and under which prompt version.
Confidence is a first-class field, not a stylistic hedge.
Peer critiques persist alongside accepted claims for review.
The full reasoning trace can be replayed and inspected after the fact.
No consequence graph is treated as final without an accountable human reader.
Severe drought · tracing 20% less rainfall through six domains.
The public prototype's flagship scenario traces the cascading effect of a sustained 20% reduction in rainfall. Six domain agents — water, agriculture, biodiversity, economy, public health, and infrastructure — inherit each other's conclusions and record where their reasoning aligns, diverges, or depends on assumptions the next agent would challenge.
Reservoirs, aquifers, and municipal supply under sustained 20% rainfall deficit.
Crop yields, planting decisions, and livestock stress under revised water budgets.
Riparian ecosystems, migratory patterns, and species stress inherited from the water and agriculture layers.
Commodity prices, employment, and regional GDP effects reasoned from upstream constraints.
Heat exposure, air quality, and food-system stress traced from environmental and economic conditions.
Grid load, cooling demand, transport, and water-treatment capacity under compounded pressure.
What has been shipped, what is scaffolding.
- July 19, 2026 · 09:00Concept framingFramed Interloom as a research question about legibility: whether cross-domain consequences become more usable when specialized agents commit to typed, inspectable claims rather than producing a single narrative answer.
- July 19, 2026 · 14:00Legible interface prototypeBuilt the first end-to-end interface around the severe-drought scenario. The public experience uses curated scenario traces for reliability and legibility, while the technical paper documents the evidence-grounded dynamic architecture that the next iteration will implement.
- July 19, 2026 · 18:00Public research releasePublished the live prototype, the technical paper, and the public repository. Interloom is released as a research prototype so that its architecture, limitations, and next experiments are open to external scrutiny from day one.
What Interloom is not.
Interloom is decision support, not a deterministic forecast. It is not an emergency alert system, a policy authority, medical advice, or a substitute for expert judgment. Every claim is a hypothesis the human reviewer is expected to interrogate. Where the current prototype uses curated traces, that limitation is stated in the interface itself.
What the paper commits to next.
- EvidenceEvidence registry & citations
A retrievable evidence store so every claim links to inspectable sources.
- ProtocolTyped claim bus
A shared schema for claims, dependencies, and uncertainty passed between agents.
- DialecticAgent critique & dissent
Structured disagreement between agents preserved alongside accepted conclusions.
- CalibrationConfidence calibration
Measuring whether stated confidence tracks empirical accuracy across scenarios.
- BenchmarksExpert benchmark scenarios
Curated scenarios reviewed by domain experts to score system-level reasoning.
- TransferPolicy, healthcare, supply chains, biology
Porting the architecture into further high-stakes interconnected domains.
The drought scenario, walked end to end.
Prefer a local copy? Download the demo (MP4).