Skip to content
AffixIO AFFIXIO
SDK docs Contact Us

Agentic systems

AI agent verification

Gate AI agents before they call tools, move funds, or take autonomous action. AffixIO evaluates policy per request and returns ML-DSA-65 signed allow or deny with Merkle audit evidence. No profile store at the verifier. Integrate via REST API, Node.js SDK, or MCP connector for Claude and Cursor.

Contact UsTry the sandboxMCP connector
KYAPOLICY GATESML-DSA-65MCPMERKLE AUDITZERO PII

At a glance

Product
Signed allow or deny before autonomous agent actions
Scope
Tool calls, payments, data retrieval, workflow steps
Signatures
ML-DSA-65 post-quantum (NIST FIPS 204)
Integration
REST API, Node.js SDK, MCP connector
Audit
Merkle-anchored proof references
Retention
Zero PII at verifier by default

AI agent verification is a policy gate that returns a signed allow or deny before an autonomous agent executes a consequential action. AffixIO separates authorisation from the agent runtime so auditors can verify what was permitted without replaying prompts or storing user profiles at the gate.

The challenge

Why agents need a verification boundary

Autonomous agents act at machine speed. Without a verification boundary, authorisation is buried in prompts, heuristics, or post-hoc logs that auditors cannot replay.

Prompt-only guardrails

LLM instructions are not cryptographic evidence. They fail under adversarial input and cannot be audited after the fact.

Scope creep

Agents that can call any tool accumulate privilege. Each integration expands blast radius without a binary gate.

Post-hoc forensics

Transaction logs show what happened, not what policy permitted at decision time.

Regulatory pressure

EU AI Act Article 12 and OWASP Agentic Top 10 expect traceable controls on autonomous actions.

Use cases

Where agent verification applies

Any autonomous action that should be explicitly permitted or denied before execution.

Tool call authorisation

Gate MCP tool invocations before data leaves the agent sandbox.

MCP · policy

Delegated payments

Verify spend limits and consent before an agent initiates payment.

AP2 · treasury

Data retrieval

Confirm the agent may access a dataset without exporting records to the gate.

ZK · ABAC

Workflow steps

Multi-step automations get a signed outcome per step.

pipeline

Human escalation

Route denied actions to review with proof metadata attached.

audit ref

Cross-agent delegation

Sub-agents inherit bounded credentials with spent-proof anti-replay.

delegation

Proof pipeline

How verification works

  1. 01

    Issue

    Policy authority defines a Noir circuit for the rule set. Witness inputs come from systems that already own the records.

  2. 02

    Bind

    Proof is bound to event context: programme ID, expiry, geography, and anti-replay nonce. ML-DSA-65 signature applied at issue.

  3. 03

    Present

    User, agent, or device presents the credential via API call or secure QR at the service boundary.

  4. 04

    Verify

    Verifier receives allow or deny plus Merkle reference. No personal data retained at the verifier by default.

Stack

Technical foundation

ML-DSA-65Post-quantum signatures
NoirZero-knowledge circuits
MerkleAudit anchoring
RESTOpenAPI 1.4.2
SDK@affix-io/sdk
QROffline spent-proof

Proof generation can run on infrastructure you control. See technical architecture and circuit catalogue.

FAQ

Frequently asked questions

What is know your agent (KYA)?

KYA is the practice of verifying an agent's authority to act before it executes. AffixIO returns a signed outcome, not a risk score.

How does AffixIO integrate with MCP?

The MCP connector exposes verification endpoints for Claude, Cursor, and compatible clients. See the MCP documentation.

Can verification run on our infrastructure?

Yes. The Node.js SDK issues and verifies proofs on infrastructure you control.

What happens on deny?

The verifier receives deny with signed proof metadata. Downstream systems block the action and may route to human review.

Does AffixIO store conversation data?

No. By default the verifier receives allow or deny and proof metadata, not prompts or user profiles.

Which regulations does this align with?

Architectural alignment with EU AI Act audit evidence patterns and OWASP Agentic Top 10. Not a certification claim.

Get started

Validate ai agent verification on the live sandbox, then scope a pilot with engineering.

Contact Us Open sandbox Start evaluation

Related resources