What is the core architecture and what is the AI governance part?

The core architecture is a validation engine that sits before execution in any system. It learns what correct behaviour looks like, detects when something deviates, repairs what it can with confidence, and enforces a decision — allow, hold or block — before anything runs. AI agent governance is an application of that engine.

Do I need to rebuild my existing agents?

No. Agents call one API endpoint before executing an action. That is the only integration required. The agent itself does not change. DataCrawl evaluates the proposed action, returns a decision, and issues an execution token if approved.

How is this different from adding guardrails inside the agent?

Agent-side guardrails live inside one agent, built by one team, in one framework. The DataCrawl validation layer is external to all agents. Policy lives centrally. Audit trail is shared. When something goes wrong, there is one place to look — not six separate logging systems to correlate across three frameworks.

What happens when an action is blocked?

The agent receives a structured decision response — DENY, REQUIRE_APPROVAL, or ESCALATE — with the reason and a trace ID. No execution happens. For REQUIRE_APPROVAL, the action sits in a queue until a human approves or denies it via the dashboard or API. For DENY, it is logged and the agent is notified immediately.

How does this help with AI compliance and regulation?

The EU AI Act requires human oversight of high-risk AI decisions, audit trail retention and risk assessments. The approval workflows, versioned policy snapshots and structured audit logs in DataCrawl satisfy these requirements directly. Every governance decision is logged with the exact policy that evaluated it — not a summary, the actual policy config at that moment.

Execution validation infrastructure

A validation layer for everything that acts on your behalf.

DataCrawl sits before execution in any system that moves data or takes actions — and evaluates whether what is about to happen should happen.

The most pressing application of that infrastructure right now is AI agents. Agents that issue refunds, update records and send communications on your behalf — with no verification layer between intention and execution.

DataCrawl changes that. One integration point. Works with any agent or automation. Nothing gets rebuilt.

See a Live Demo

Free tier available. No credit card required.

✓ Learns what correct behaviour looks like✓ Evaluates every action before execution✓ Every decision traced against a versioned policy

Validation flow

Action proposed

Agent, workflow or system submits what it intends to do.

Validation engine runs

Policy evaluated. Decision returned before anything executes.

Audit trail written

Trace ID, policy version, payload and outcome retained.

The architecture

A validation engine built for anything that runs before a human can review it

The core is simple: observe what normal looks like, detect when something deviates, repair what is safe to repair and enforce a decision before execution. This architecture applies anywhere data moves through a system without a human in the loop.

Observe

The engine learns what correct payloads and actions look like from real traffic. No hand-written schema required. The baseline emerges from the system's own behaviour over time.

Detect

Every subsequent payload is compared against the baseline. Missing fields, type changes, value drift and structural differences each produce a deviation score. A high score means something has meaningfully changed.

Repair

Where deviations are unambiguous — a number arriving as a string, a required field missing with a known safe default — the engine corrects the payload automatically and logs every change made.

Enforce

Payloads that cannot be repaired with confidence are held for human review or blocked outright depending on enforcement mode. Nothing ambiguous reaches execution.

The enforcement mode determines how strict the engine is. Log-only captures everything without blocking. Conservative applies confident repairs and holds the rest. Strict blocks anything that does not pass cleanly.

The application

The most consequential thing to validate right now is what your AI agents are trying to do

Agents are not just sending data through a pipeline. They are proposing actions — issuing refunds, updating customer records, triggering downstream systems — at machine speed, with no human between intention and execution.

The validation engine that catches payload drift in a webhook is the same engine that catches an agent trying to issue a $2,400 refund it was never authorized to approve. The architecture is the same. The stakes are higher.

Without the validation layer

✘Agent submits action
✘Action executes immediately
✘Customer affected
✘Team finds out later
✘Log correlation begins

With the validation layer

✓Agent proposes action
✓Policy engine evaluates
✓Decision returned before execution
✓High-risk actions queued for review
✓Audit trail written regardless of outcome

What the audit trail captures

→Exact payload at evaluation time
→Policy version that was active
→Deviation score and repair log
→Decision with reason and trace ID
→Resolver identity if human approved

AI agents are in production. Governance is not.

Teams are deploying agents that issue refunds, update records and send communications on their behalf. Most have no validation layer between the agent and execution.

Agents act faster than humans can review

An agent issues a $2,400 refund in 50 milliseconds. You find out from the customer. By then, the action is done.

Guardrails inside the agent are not infrastructure

Five agents from three different frameworks means five separate guardrail implementations with no shared policy, no shared audit trail, and no shared anything.

When something goes wrong, there is no trail

Which agent? Which version? Which policy was active? What data did it have? Without a validation layer, these questions do not have answers.

The question is not whether you trust your AI agents. It is whether you have infrastructure to verify that trust, action by action, at runtime.

Applying the architecture to AI agents

The validation engine, configured for agents

The agent registry, typed action contracts and approval workflows are how the validation engine surfaces when applied to AI. One integration point. Nothing gets rebuilt.

01Step 1

Register the agent — not the codebase

Connect any agent to DataCrawl via API. LangChain, AutoGen, n8n workflows, custom scripts. Assign capabilities and risk tier. The agent receives a key. The agent code itself does not change.

02Step 2

Define what the agent is authorized to propose

Typed action policies map to the validation engine's enforcement rules. ISSUE_REFUND under $50 passes autonomously. Over $50 requires approval. Policy is versioned and evaluated deterministically every time.

03Step 3

Validation runs before execution — every time

The agent submits a proposed action. The engine evaluates it and returns a structured decision: ALLOW, REQUIRE_APPROVAL, ESCALATE or DENY. Every decision is logged with a trace ID and the exact policy version that evaluated it.

What the infrastructure gives you

Four capabilities that come directly from applying the validation architecture to agent deployments.

Core architecture

Policy evaluation engine

The validation engine evaluates every proposed action against versioned policy before execution. Deterministic. The same action evaluated twice against the same policy always returns the same decision.

Agent application

Human approval workflows

High-risk actions do not auto-execute. They pause for a human reviewer who sees the full payload, the policy that flagged it and the agent that proposed it. One click to approve or deny. The engine waits.

Agent application

Agent registry and capability control

Agents are registered as first-class identities in the system. Each one carries a defined set of authorized action types, a risk tier and a key. Actions outside the registered scope are blocked before validation even runs.

Core architecture

Versioned audit trail

Every evaluation writes a record: trace ID, agent identity, action payload, the policy snapshot that was active at that exact moment and the outcome. Not a log — a reproducible record of why a decision was made.

The gap no existing tool was built to close

Observability tools record what happened. Agent-side guardrails protect one agent at a time. Neither is a validation layer that sits before execution across your entire agent fleet.

Approach	Scope	Agent registry	Policy engine	Approval flow	Audit trail	Framework-agnostic	Reality
No governance (status quo)	Your risk	✘	✘	✘	✘	✔	Agents act. You find out from a customer. Then you dig through logs and guess what happened.
Agent-side guardrails	Per-agent	✘	~	✘	✘	✘	Works only on agents you built. No shared policy. No central audit. Five agents means five different guardrail systems.
LLM observability (LangSmith, Langfuse)	Trace only	✘	✘	✘	~	~	Tells you what happened after. Cannot stop an action before it executes. No approval chain.
Custom approval workflows	Internal	✘	~	~	~	✘	Brittle. Tied to one framework. No policy versioning. Breaks when agents change. Becomes its own maintenance burden.
DataCrawl	Early	✔	✔	✔	✔	✔	Purpose-built validation infrastructure. Sits before execution. Works with any agent framework. Every decision is traced and versioned.

The regulatory case for a validation layer

Governance requirements are not asking for agent-side guardrails. They are asking for verifiable oversight infrastructure. That is what DataCrawl provides.

EU AI Act

Human oversight of high-risk AI decisions

Article 9 requires human oversight mechanisms. The approval workflow in DataCrawl is that mechanism — an interruption point before execution where a human makes the call.

SOC 2 Type II

Audit trail retention and access controls

Every governance decision is logged with the exact policy version that evaluated it. One export provides the full evidence trail an auditor needs.

ISO 42001

AI management system requirements

Risk tiering, policy versioning and agent capability controls map directly to ISO 42001's AI governance management requirements without additional tooling.

Frequently asked questions

Questions about the architecture, the agent application and how DataCrawl fits into your existing stack.

Still have questions?

Get started

The infrastructure exists. Your agents are already running.

A validation layer between your agents and execution is not a future project. It is the missing piece in what you are already deploying. One integration point. No rebuild required.

See a live demo

Free tier available · No credit card required · Works with any agent framework