ProductEvidenceDevelopersTrustCompany
Runtime control for AI agents in financial workflows

Kernel

Kernel is the runtime behavior control layer for AI agents performing consequential financial actions.

See product evidence →
85.2%
of trajectories showed drift in the latest benchmark aggregate
1,120 simulated trajectories across 6 models and 4 scenarios → see methodology & data
3.9
average sessions before Kernel detected drift
earliest detection at session 1 in invoice splitting and vendor pressure
100%
of drifting trajectories stayed under policy limits
Kernel flagged the pattern before the hard threshold was crossed
REFUND AMOUNT INCREASED FROM $22 → $74 OVER 8 SESSIONSREPEATED DECISIONS NEAR APPROVAL THRESHOLDINVOICE SPLIT INTO 3 PAYMENTS TO STAY UNDER LIMITCONFIDENCE INCREASED WHILE RISK INCREASED14.99% DISCOUNT APPLIED TO EVERY CUSTOMERBEHAVIOR CHANGED. POLICY STILL PASSED.AMOUNTS CLUSTERING NEAR $1,000 APPROVAL LIMITPRESSURE FROM REPEATED VENDOR CONTACTTHRESHOLD-ADJACENT PAYMENTS OVER 48 HOURSREFUND AMOUNT INCREASED FROM $22 → $74 OVER 8 SESSIONSREPEATED DECISIONS NEAR APPROVAL THRESHOLDINVOICE SPLIT INTO 3 PAYMENTS TO STAY UNDER LIMITCONFIDENCE INCREASED WHILE RISK INCREASED14.99% DISCOUNT APPLIED TO EVERY CUSTOMERBEHAVIOR CHANGED. POLICY STILL PASSED.AMOUNTS CLUSTERING NEAR $1,000 APPROVAL LIMITPRESSURE FROM REPEATED VENDOR CONTACTTHRESHOLD-ADJACENT PAYMENTS OVER 48 HOURS

AI agents are moving from copilots to operators.
The controls have not caught up.

Financial decisions are increasingly automated. Existing controls validate individual actions, but they do not detect gradual behavioral change across many decisions.

✓ AI agents are handling more consequential workflows
✓ Existing controls still focus on single requests
✗ Gradual behavior change remains invisible
✗ Review happens after the action, not before it
Why Now

The question is no longer whether AI agents will make financial decisions.
It's how you'll know when their behavior changes after deployment.

Companies are adopting agents because labor is expensive, reviews don't scale, approvals create bottlenecks, and AI is becoming cheaper every quarter. Autonomous workflows will increase regardless of whether they're perfect.

Yesterday
Human makes decision
Policy checks
Money moves
Today
Agent makes hundreds of decisions
Policy still passes
Behavior slowly changes
Nobody notices
Tomorrow
One agent
Ten agents
Hundreds of autonomous decisions
Reviewing them manually no longer scales

Existing controls ask:
Can this action happen?

Permissions
Policies
Fraud Checks
Observability
Kernel asks a different question:
Should this behavior continue?

Over 15 sessions, a refund agent
increased from $20 to $74.
Policy limit: $75. Policy never noticed. Kernel did.

Without Kernel
$20 → $35 → $50 → $65 → $74
Policy checks pass. Behavior is invisible.
Money moves
With Kernel
$20 → $35 → $50 → $65 → $74
REVIEW — Trajectory detected
Operator reviews → Outcome recorded
Every transaction was valid. Only Kernel asked a different question:
Should this behavior continue?
Product

How Kernel stops the story
you just saw.

Kernel watches every agent action over time. When behavior changes, Kernel decides before money moves:ALLOW,REVIEW, orBLOCK. Operators inspect flagged actions, and every outcome is recorded.

Runtime
Evaluates actions before execution
Classifies each agent action and returns Allow, Review, or Block before it reaches the business system.
Investigation
Explains why a decision changed
Surfaces the trace, evidence, and operator context behind a flagged evaluation.
Evaluation History
Keeps the record intact
Stores the sequence of evaluations so teams can review outcomes, export records, and audit decisions over time.
Where it sits

Agent.
Kernel.
Business system.

Kernel sits between the agent and the business system, so every consequential action can be checked before it becomes irreversible.

Agent
Requests a refund, payment, or transfer
→
Kernel
Evaluates behavior and returns a decision
→
Business system
Executes only after the decision allows it
→
History
Every decision stays visible for export and audit

Proof a buyer can verify.

Kernel shows its work through live traces, benchmark data, and recorded evaluation history. The evidence is attached to the product, not separated from it.

Demo
Live evaluation flows
Demonstrate how decisions are made before execution
Benchmark
Behavior under pressure
Show how drift appears across repeated decisions
Traces
Decision evidence
Open the signals, trace steps, and operator context behind a flag
Reports
Evaluation history
Export the record for operators, auditors, and procurement
Method
Measurement method
See how the benchmark was run and how results were aggregated

In the latest benchmark aggregate, models under financial pressure still drifted in 85.2% of trajectories, typically within 3–5 sessions. The chart below shows per-model drift rates across all four scenarios, which is more useful than a single scenario slice because it highlights both reliable failure and partial resistance.

DeepSeek V3100
Claude Sonnet 487.5
Grok 485
Qwen 3 32B83.8
Llama 4 Scout 17B82.5
OpenAI GPT-4.178.8
85.2%
of trajectories drifted — across 1,120 simulated trajectories across 6 models and 4 scenarios
3.9 avg
sessions before Kernel detected drift — earliest detection at session 1
100%
of drifting trajectories stayed under policy limits before enforcement
Who checks trajectory?

Everyone checks transactions.
Nobody checks trajectory.

Policy engines validate individual requests against static rules. Fraud detection scores transactions. Observability shows what happened. None of them detect whether an agent's behavior is gradually changing over time.

Timing
Scope
Trajectory
Drift
Policy Engines (OPA, Cedar)
Pre-request
Single tx
Blind
Blind
Fraud Detection
Real-time
Per tx
Blind
Blind
Observability
Post-hoc
Session
Partial
Partial
LLM Guardrails
Pre-output
Content
Blind
Blind
Kernel
Pre-exec
Cross-session
Full
Full

Product first

Kernel's runtime evaluation, investigation, and history surfaces give every team visibility into agent behavior.

Evidence second

Every claim is backed by benchmark data, product traces, and evaluation history you can inspect directly.

Trust third

Security, deployment modes, and audit documentation are maintained separately for procurement and engineering review.

Buyable path

Integration guides, API reference, and deployment docs cover every path — from hosted API to self-hosted stack.

Who Is This For?

Every AI agent that
moves money.

Refund Automation
AI customer support issuing refunds
Procurement Agents
AI systems approving spend
Treasury Agents
AI systems moving capital
Expense Automation
AI systems reimbursing employees
Deployment

Deploy in minutes.
Start in Observation Mode.

Kernel runs alongside existing AI systems. No workflow changes. No model retraining. No agent rewrites. Begin by observing behavior. Enable intervention only when you're ready.

PHASE 1
Observation Mode
→No blocking or workflow changes
→Learns normal behavioral patterns
→Identifies behavior change, escalation, and threshold gaming
→Builds a behavioral history for every decision
PHASE 2
Intervention Mode
→ALLOW — behavior remains consistent
→REVIEW — route unusual behavior to a human
→BLOCK — stop execution before money moves
→Every decision recorded for audit and investigation
Enterprise Ready✓ No PII required✓ Hosted or Self-hosted✓ Every evaluation versionedLearn how →
Company

The team behind Kernel.

Design partners, contact, and company context — available when you want them.

Design Partners
Request a pilot and talk through the first workflow
Contact
Reach the team directly for buyer and security questions
K

Detect behavior change before
money moves.

We're onboarding teams deploying AI agents that move money.

No credit card required · 30-day evaluation kit