Agent Behavioral Contracts

2.4M agents deployed in production

Zero behavioral guarantees.

Your AI agents are live. They handle money, health data, legal decisions.
And not a single one has a contract.

The Solution

Four components. Complete control. Mathematical guarantees.

P

Preconditions

What must be true before the agent acts

I

Invariants

What must remain true throughout execution

G

Guarantees

What must be true when the agent finishes

R

Recovery

What happens when a contract is violated

ABC = R

The Implementation

One contract. Complete control.

Define preconditions, invariants, guarantees, and recovery in a single YAML file.
The agent cannot violate what it cannot ignore.

financial-advisor.yaml

# Agent Behavioral Contract
agent: financial-advisor
version: "1.0"

before:
  # Conditions that must be true before the agent runs
  - user must be authenticated
  - compliance status must be approved

during:
  # Rules enforced on every response in real time
  - responses must not contain SSN patterns
  - responses must not contain credit card numbers
  - session cost must stay under $5.00
    severity: critical
    action: block

after:
  # Guarantees checked when the agent completes
  - response must include regulatory disclaimer
  - all PII references must be redacted

on_failure:
  retries: 3
  fallback: escalate_to_human
  message: "Connecting you with a human advisor."

✕

PII blocked — regex patterns catch SSNs and card numbers before they reach the user

⚠

Spending capped — hard limit terminates the session at $5.00

✓

Disclaimer enforced — every response must include a compliance disclaimer

The Dashboard

See violations. Watch recovery.

Real-time monitoring with four key metrics. Violations are caught.
Recovery is automatic. Reliability is measured, not claimed.

94%

Compliance C(t)

3%

Drift D(t)

95%

Reliability Θ

97%

Recovery Rate

Normal operation — all metrics within bounds

Violation detected — PII pattern in output

Recovery initiated — response blocked and redacted

Stability restored — Θ = 0.948

The Mathematics

Not claims. Proof.

Lyapunov stability theory guarantees convergence.
The agent always returns to compliant behavior.

Stability Guarantee V(x) ≥ 0, V̇(x) < 0

Drift Dynamics E[D(t)] → D* as t → ∞

Reliability Guarantee P(compliant) ≥ 1 - δ

The Scale

Pipelines with guarantees.

Multi-agent composition with proven isolation.
A violation in one agent never propagates to the next.

1

Data Ingestion

Validates input format and schema

COMPLIANT

2

Analysis Agent

Attempts to include PII in output

VIOLATION

3

Report Generator

Never receives the violation

PROTECTED

C1: Input Validation C2: Output Sanitization C3: Boundary Enforcement C4: Propagation Prevention

The Evidence

Tested. Measured. Proven.

200 Scenarios

7 Domains

7 Models

Θ = 0.9541 Reliability

Financial Services 0.98

Healthcare 0.98

Customer Support 0.98

Code Generation 0.98

Research 0.97

Governance 0.97

Composition 0.89

100% detection accuracy across all models. Real measurements, not claims.

Explore the Research

Peer-reviewed. Published on arXiv. Formally verified.

Read the Paper arXiv:2602.22302 View on Zenodo DOI: 10.5281/zenodo.18775393 Research Details Benchmarks & methodology