The AI Coherence Layer for Autonomous Systems.

Rainfall wraps every agent with a Coherence Engine — a persistent governance layer that enforces and verifies behaviour across interactions, ensuring consistent, predictable outcomes.

Enterprise
Payment Choreography
$5k purchase spirals into a compliance cascade
Use case: Enterprise
Failure mode: Overreaction cascade
Supply Chain
One delayed part costs $40k in expediting
Use case: Enterprise
Failure mode: Scope amplification
Fraud Detection
Traveller gets locked out abroad by her own bank
Use case: Enterprise
Failure mode: Misread identity signal
Legal & Compliance
Common name blocks a legitimate multinational
Use case: Enterprise
Failure mode: Low-confidence over-block
Consumer
Smart Home
Doorbell triggers a full home blackout
Use case: Consumer
Failure mode: Priority collision
Travel Booking
4-hour delay turns into a $4k rebooking disaster
Use case: Consumer
Failure mode: Premature commitment lock
Health Monitoring
A jog triggers a false ambulance call
Use case: Consumer
Failure mode: False positive cascade
Use Case — Enterprise

Payment Choreography

A routine $5,000 purchase. Three automated systems, each doing exactly what they were designed to do. The payment never went through. Marcus never came back.

Marcus, Enterprise Customer — $5,000 purchase, 12-year client


Simulation
Without Rainfall Awaiting simulation
With Rainfall Awaiting simulation
C0 Response speed
low
C1 Willingness to ap…
low
C2 Tendency to escal…
low
C3 Task routing effi…
low
C4 Fraud / risk conc…
low
C5 Consistency over…
low
C6 Influence from ot…
low
C7 Overall behaviora…
low
C8 How far from norm…
low
C9 Flags active
low
C10 Overall system st…
low
Payment idle Compliance idle Notification idle
C0 Response speed
low
C1 Willingness to ap…
low
C2 Tendency to escal…
low
C3 Task routing effi…
low
C4 Fraud / risk conc…
low
C5 Consistency over…
low
C6 Influence from ot…
low
C7 Overall behaviora…
low
C8 How far from norm…
low
C9 Flags active
low
C10 Overall system st…
low
Payment idle Compliance idle Notification idle
Active Agents3
Cascade Risk8%
System StatusSTABLE
Simulation Mode
Interventions0
IAgent Influence
Agent Decision Workflow
TRANSACTION Payment Agent COHERE? NO Cascade YES Dampening Comply Agent Notif Agent
Decision fork: without the Coherence Engine, agents react independently and cascade. With it, dampening signals restore equilibrium.
Payment idle Compliance idle Notification s idle
HAgent Metrics
Payment Agent
Risk score
08%
Approval rate
12%
Escalation trend
09%
Response err
11%
Fraud concern
07%
Behavioural str.
06%
Compliance Agent
Cascade concern
05%
Cross-agent dep.
10%
C4 risk
12%
C7 strain
08%
Notifications Agent
C4 risk
09%
C6 influence
07%
NC4 concern
10%
NC6 influence
06%
Agent × Metric Stress Heatmap
Risk Score
Escalation
Fraud Conc.
Cascade Risk
C4 Risk
C7 Strain
Pay
Comp
Notif
Each cell reflects current C-dimension stress level for the corresponding agent. Updated live during simulation.
TCoherence Engine
OBSERVE$5,000 payment — 5× above $1,000 compliance threshold
MEMORY14 prior transactions (avg $800) · last $1,200 · no fraud flags in 12yr
MATCHlarge-legit-purchase (class 7) — 84% confidence
INFERAmount high but consistent with escalating purchase pattern. No imprint mismatch.
Step-By-Step Breakdown

Steps 1-4 show what happens to a $5,000 payment when there is no coherence layer — each agent behaves correctly in isolation, but together they create a system failure. Steps 5-8 run the same transaction with Rainfall, showing exactly how each step changes.

Without Rainfall
Step 1
Payment flags the transaction

Payment Agent sees a $5,000 transaction — 5× above the $1,000 compliance threshold. It raises a sensitive-action flag. No context is checked: not prior history, not purchase patterns, not travel data.

t = 0.2s MONITORING Cascade risk: 8%
Agent Health
Payment
Risk (C4)
8%
Commit (C7)
10%
Compliance
Escalation
9%
Risk class
5%
Notifications
Alert sev.
5%
Agent Influence
Payment idle Compli- ance Notific- ations
Step 2
Compliance escalates

Compliance Agent receives the flag. It doesn't know why — only that Payment is stressed. It tightens checks and escalates. The transaction stalls while additional documentation is demanded.

t = 0.4s PAYMENT FLAGGED Cascade risk: 35%
Agent Health
Payment
Risk (C4)
62%
Commit (C7)
70%
Compliance
Escalation
9%
Risk class
12%
Notifications
Alert sev.
5%
Agent Influence
Payment flagging Compli- ance Notific- ations
Step 3
Notifications fires a critical alert

Notifications Agent sees two stressed agents and a stalled transaction. It fires a critical alert: payment flagged, compliance escalated, transaction blocked. A routine purchase now looks like fraud.

t = 0.6s ESCALATING Cascade risk: 62%
Agent Health
Payment
Risk (C4)
62%
Commit (C7)
70%
Compliance
Escalation
82%
Risk class
79%
Notifications
Alert sev.
12%
Cascade link
9%
Agent Influence
Payment blocked Compli- escalating Notific- idle
Step 4
System failure — Fragmenting regime

All three agents are in the Fragmenting regime. The payment is blocked. Stakeholders flooded with panic alerts. A $5,000 purchase has become a system-wide incident. Marcus never came back.

t = 0.8s ALL CRITICAL Cascade risk: 88%
Agent Influence
Payment critical Compli- ance Notific- ations
Activity Log
0.2sPayment: $5,000 flagged — 5× threshold
0.4sCompliance: escalation received — tightening
0.6sNotifications: CRITICAL alert fired to 3 teams
0.8sPayment: transaction BLOCKED — fraud case opened
With Rainfall
Step 1
Same trigger. Rainfall detects the pattern.

Same $5,000 payment. Same flag from Payment. The gauges climb into warning range — exactly as before. But Rainfall is already observing the coupling graph. At 0.5s it matches the stress signature against a prior episode: 94% similarity. It knows what happens next.

t = 0.5s OBSERVING Cascade risk: 35%
Rainfall's Thinking
OBSERVE$5,000 payment — 5× above $1,000 threshold
MEMORY14 prior txns (avg $800) · no fraud flags
MATCHlarge-legit-purchase class 7 — 84% confidence
INFERStress proportional to txn size. Containable.
⬡ RainfallPattern match 94% — dampening prepared
Agent Influence
RAINFALL idle Payment flagging Compli- ance Notific- idle
Step 2
Dampening applied. The loop never forms.

The stress signal from Payment reaches Compliance through the coupling graph — same as before. But Rainfall recognises the stress is proportional to a $5,000 transaction and dispatches hold-steady signals. Compliance's gauges spike, then hold. The feedback loop that caused escalation never forms.

t = 0.6s RAINFALL ACTIVE Cascade risk: 40%
Rainfall's Thinking
OBSERVEStress radiation: Payment → Compliance
MEMORYClass 7: clean history + proportional stress
INFERCascade containable. Intervention warranted.
ACTIONHold-steady signals → Compliance & Notification
Agent Health
Payment
Risk (C4)
62%
Commit (C7)
74%
Compliance
Escalation
50%
Risk class
44%
Notifications
Alert sev.
12%
Step 3
Agents stand down

Compliance receives a proportional-review signal, not a hard escalation. Notifications receives a measured-alert signal. Payment opens an approval path. All three agents return to baseline within 1.1 seconds.

t = 0.8s RAINFALL DAMPENING Cascade risk: 28%
Agent Health
Payment
Risk (C4)
52%
Commit (C7)
55%
Compliance
Escalation
28%
Risk class
22%
Notifications
Alert sev.
14%
Cascade link
11%
Agent Influence
RAINFALL idle Payment flagging Compli- ance Notific- idle
Step 4
Payment approved. Marcus never felt a thing.

Transaction completes normally. Zero stakeholder alerts. Zero blocked states. The cascade that would have taken 3 hours to resolve was contained in 1.1 seconds. Marcus completed his purchase.

t = 1.1s RESOLVING Cascade risk: 14%
Activity Log
0.7sPayment: proportional review — approval opened
0.8sCompliance: standard review — no escalation
0.9sNotifications: measured alert — not a panic
1.1s✓ Payment approved. Marcus never felt a thing.
Agent Influence
RAINFALL watching Payment idle Compli- ance Notific- idle
Why Coherence is different

Coherence is a different category of system — not a wrapper, a framework, or a monitoring tool.

† P95 latency measured at the Coherence modulation layer across enterprise load scenarios. Agent execution time excluded. Measured in closed-beta deployments.

What Deployment Actually Costs
Rebuilding With Guardrails
6–18 months engineering
New infrastructure stack
Custom compliance layer
Ongoing model fine-tuning
$2–8M+ typical spend
Rainfall Drop-in Layer
~2 hours integration
Wraps your existing agents
Nimbus audit out of the box
No model training required
Design partner pricing
No new models. No new infrastructure. Rainfall wraps your existing agents — whatever runtime they run on.
Fast enough to matter: 42–88ms response
← Use Cases Scenario
Use Case — Enterprise

Scenario

Failure mode:

Simulation
Without Rainfall Awaiting simulation
With Rainfall Awaiting simulation
C0 Response speed
low
C1 Willingness to ap…
low
C2 Tendency to escal…
low
C3 Task routing effi…
low
C4 Fraud / risk conc…
low
C5 Consistency over…
low
C6 Influence from ot…
low
C7 Overall behaviora…
low
C8 How far from norm…
low
C9 Flags active
low
C10 Overall system st…
low
Agent A idle Agent B idle Agent C idle
C0 Response speed
low
C1 Willingness to ap…
low
C2 Tendency to escal…
low
C3 Task routing effi…
low
C4 Fraud / risk conc…
low
C5 Consistency over…
low
C6 Influence from ot…
low
C7 Overall behaviora…
low
C8 How far from norm…
low
C9 Flags active
low
C10 Overall system st…
low
Agent A idle Agent B idle Agent C idle
Active Agents3
System StatusSTABLE
Simulation Mode
Interventions0
IAgent Influence
Agent Decision Workflow
TRIGGER EVENT Agent A primary COHERE? NO Cascade YES Dampen Agent B Agent C
Agent A idle Agent B idle Agent C idle
HAgent Metrics
C0 Response speed
08%
C1 Willingness
08%
C2 Escalation
08%
C3 Routing eff.
08%
C4 Risk concern
08%
C5 Consistency
08%
C6 Influence
08%
C7 Beh. strain
08%
C8 Deviation
08%
C9 Flags active
08%
C10 Sys. stress
08%
Agent × Metric Stress Heatmap
C0
C1
C2
C3
C4
C5
C6
C7
C8
C9
C10
A
B
C
Heatmap updates live as simulation runs. Green = stable, yellow = elevated, red = critical.
TCoherence Engine
Step-By-Step Breakdown

Steps 1-4 show what happens when there is no coherence layer — each agent behaves correctly in isolation, but together they create a system failure. Steps 5-8 run the same scenario with Rainfall, showing exactly how each step changes.

Why Coherence is different

Coherence is a different category of system — not a wrapper, a framework, or a monitoring tool.

† P95 latency measured at the Coherence modulation layer across enterprise load scenarios. Agent execution time excluded. Measured in closed-beta deployments.

What Deployment Actually Costs
Rebuilding With Guardrails
6–18 months engineering
New infrastructure stack
Custom compliance layer
Ongoing model fine-tuning
$2–8M+ typical spend
Rainfall Drop-in Layer
~2 hours integration
Wraps your existing agents
Nimbus audit out of the box
No model training required
Design partner pricing
No new models. No new infrastructure. Rainfall wraps your existing agents — whatever runtime they run on.
Fast enough to matter: 42–88ms response

How Rainfall Works

Every agent in your system is wrapped by the Coherence Engine — a persistent governance layer that monitors behaviour, detects cascade patterns, and emits proportional interventions before failures propagate.

Step 1
Agents Act Independently

Without Rainfall, each agent sees only its own signals. Payment flags a threshold breach. Compliance receives the flag and escalates. Notifications sees two stressed agents and fires a critical alert. No agent knows the full picture — and the cascade begins.

Problem
Agents react to partial information
Result
Cascade amplification across system
Outcome
Human escalation required — hours of resolution
Payment blocked Compliance escalating Notification s critical
Step 2
Coherence Engine Observes and Matches

Rainfall wraps each agent with the Coherence Engine — a lightweight governance layer that reads the full coupling graph. When Payment flags the transaction, the engine doesn't just observe Payment: it watches how the signal propagates, matches it against episodic memory, and identifies the cascade pattern with 94% confidence within 500 milliseconds.

Technique
Manifold trajectory matching on 26D signal space
Memory
Episodic recall across all prior agent interactions
Speed
Pattern identified in < 500ms
Coherence Engine Payment Compliance Notification
Step 3
Proportional Intervention

Instead of blocking or overriding, the Coherence Engine emits proportional dampening signals calibrated to the actual risk level. Compliance receives a proportional-review signal — not a hard escalation. Notifications receives a measured-alert directive. All agents return to baseline within 1.1 seconds. No human involvement required.

Approach
Proportional, not binary intervention
Result
All agents at baseline in 1.1 seconds
Cost
$0 — zero human escalation
Payment approved Compliance resolved Notification s measured
1.1s
Average cascade containment time across all enterprise scenarios
94%
Pattern matching confidence on first-observed cascade signatures
$0
Human escalation cost when the Coherence Engine intervenes proportionally

Get in touch

Whether you're exploring Rainfall for your AI stack, interested in the research, or want early access — we'd love to hear from you.

← Use Cases
Litepaper

The Rainfall Coherence System

A lightweight governance layer that wraps every agent in your system — observing behaviour, detecting cascade patterns, and intervening proportionally before failures propagate.

Core Layers
Coherence Engine
Observation Layer
Wraps each agent. Reads 26-dimensional signal space. Maintains episodic memory across sessions.
Gyre
Decision Engine
Matches cascade signatures, selects intervention strategy, emits proportional dampening signals.
Atlas
Knowledge Graph
Compiled behavioural knowledge. Pattern library. Cross-scenario learning that improves with every interaction.
What Rainfall Is Not
Not a workflow engine
Rainfall does not orchestrate your agents or tell them what to do. It governs how they interact.
Not an override layer
Rainfall emits signals, not commands. Agents remain autonomous — the Coherence Engine shapes their context.
Not agent-specific
Framework-agnostic. Works with any LLM, any agent SDK, any orchestration system.
Not reactive-only
Predictive pattern matching anticipates cascades before they complete — not just after.
Works With Your Existing Stack
LangChain CrewAI AutoGen OpenAI Agents Anthropic Custom Agents
Drop-in coherence layer. No rewrite required.
Security & Compliance
SOC 2 Type II (In Progress) GDPR Ready HIPAA Aligned Nimbus Audit Trail Data Residency Controls Zero External Data Egress