Guardrails

Hard constraints.
On every agent.

Your AI agents have enterprise data access and machine-speed decision authority. Guardrails enforce governance-defined, deterministic constraints on what agents can access, decide, and produce — not soft filters.

Deterministic constraintsGovernance-authored3 guardrail layersHITL escalationEvery event logged
CredXplain — Guardrail Status
Input: Quality ≥ 90%
Source authorised · Schema valid · Fresh < 24h
Pass
Operational: Value ₹24.5L
> ₹20L threshold → HITL escalation triggered
Escalated
Operational: Confidence 0.72
< 0.85 threshold → additional review required
Check
Output: No PII exposed
Format valid · Disclosure included
Pass

Why this matters.
Right now.

Without governed guardrails, these problems compound silently with every new data source and AI deployment.

Soft guardrails can be bypassed.

LLM-based safety checks are probabilistic. An agent chaining tools can route around pattern-based controls — creating a false sense of constraint.

Developer-defined boundaries are invisible.

Value thresholds hard-coded in config files. Access scopes set at deployment. Compliance teams cannot see or modify them. Risk teams are excluded.

Regulators ask: what prevents agents from exceeding their mandate?

Enterprises that cannot demonstrate deterministic, governance-defined boundaries face compounding regulatory risk with every new agent deployed.

From input to
governed output.

1

Define

Compliance teams author guardrail policies — input thresholds, operational limits, output constraints

2

Deploy

Guardrails applied to every agent in the mesh — inherited from governance, not configured per deployment

3

Enforce

Every agent action evaluated against guardrails before execution — deterministic, not probabilistic

4

Escalate

Threshold breaches trigger HITL routing or automatic rejection — configurable response per guardrail type

5

Log

Every guardrail event logged — check, outcome, threshold, agent identity, timestamp

What Guardrails
delivers.

🏗️

Deterministic enforcement.

Hard constraints — not probabilistic filters. Agents cannot route around governance-authored guardrails regardless of instruction or tool chaining.

📋

Governance-authored rules.

Input guardrails, operational guardrails, output guardrails — authored by compliance and risk teams, not engineering. Visible and modifiable by governance.

⚖️

Three guardrail layers.

Input gates control what data agents can consume. Operational gates govern what decisions they can make. Output gates govern what they can produce.

📊

Guardrail event logging.

Every guardrail check logged — pass, fail, threshold, trigger — for every agent action. Evidence that constraints are enforced, every time.

See Guardrails
in action.

See how Guardrails works within the Tantor governed intelligence platform.