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.
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 = 0.2sMONITORINGCascade risk: 8%
Agent Health
Payment
Risk (C4)
8%
Commit (C7)
10%
Compliance
Escalation
9%
Risk class
5%
Notifications
Alert sev.
5%
Agent Influence
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 = 0.4sPAYMENT FLAGGEDCascade risk: 35%
Agent Health
Payment
Risk (C4)
62%
Commit (C7)
70%
Compliance
Escalation
9%
Risk class
12%
Notifications
Alert sev.
5%
Agent Influence
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 = 0.6sESCALATINGCascade 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
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 = 0.8sALL CRITICALCascade risk: 88%
Agent Influence
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 = 0.5sOBSERVINGCascade 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
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 = 0.6sRAINFALL ACTIVECascade risk: 40%
Rainfall's Thinking
OBSERVEStress radiation: Payment → Compliance
MEMORYClass 7: clean history + proportional stress
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 = 0.8sRAINFALL DAMPENINGCascade 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
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 = 1.1sRESOLVINGCascade 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
Why Coherence is different
Coherence is a different category of system — not a wrapper, a framework, or a monitoring tool.
Not an LLM wrapper. Deterministic, not probabilistic. Measures 26 behavioral dimensions directly and emits dampening signals in real time. Same input, same output, every time — no hallucination, no token cost.
Not an orchestration framework. LangChain, CrewAI, and AutoGen route tasks between agents. They don't monitor how agents behave while executing. Coherence watches behavioral state continuously and intervenes before cascades form.
Not an observability tool. Datadog and Prometheus tell you what happened after the fact. Coherence acts during the event, using longitudinal behavioral memory to distinguish normal stress from a cascade in progress.
Fast enough to matter. 42–88ms tick-level response.† By the time an alert fires, the cascade has already happened. Coherence prevents it from starting.
† 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 RainfallAwaiting simulation
With RainfallAwaiting 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
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
Active Agents3
System StatusSTABLE
Simulation Mode—
Interventions0
IAgent Influence
Agent Decision Workflow
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
Episodic Memory Trace
t-5Prior episode
stored — baseline nominal
t-2Coupling graph
snapshot — edge weights updated
t-0Manifold state
recorded — 26D vector indexed
Inference Network
INPUT26D manifold — agents at
baseline. Coupling: low.
L1Episode matching — scanning 26D
episodic store
OUTAwaiting cascade signal —
monitoring active
ICoherence Insights
OBSERVEMonitoring
cross-agent stress signals...
COUPLINGGraph
topology: Stable
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.
Not an LLM wrapper. Deterministic, not probabilistic. Measures 26 behavioral dimensions directly and emits dampening signals in real time. Same input, same output, every time — no hallucination, no token cost.
Not an orchestration framework. LangChain, CrewAI, and AutoGen route tasks between agents. They don't monitor how agents behave while executing. Coherence watches behavioral state continuously and intervenes before cascades form.
Not an observability tool. Datadog and Prometheus tell you what happened after the fact. Coherence acts during the event, using longitudinal behavioral memory to distinguish normal stress from a cascade in progress.
Fast enough to matter. 42–88ms tick-level response.† By the time an alert fires, the cascade has already happened. Coherence prevents it from starting.
† 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
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
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
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.
✓
Message sent
Thank you for reaching out. A member of the Rainfall team will be in touch with you shortly.
← 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.