Interface: ICrossAgentGuardrailService
Defined in: packages/agentos/src/safety/guardrails/ICrossAgentGuardrailService.ts:127
Cross-agent guardrail service for multi-agent supervision.
Extends IGuardrailService to enable observation and intervention in other agents' output streams. Use this for:
- Supervisor patterns: A supervisor agent monitors worker agents
- Quality gates: Enforce quality standards across an agency
- Policy enforcement: Apply organization-wide rules to all agents
- Safety monitoring: Detect and prevent harmful outputs from any agent
Examples
class QualityGateGuardrail implements ICrossAgentGuardrailService {
// Observe all agents in the agency
observeAgentIds = [];
canInterruptOthers = true;
async evaluateCrossAgentOutput({ sourceAgentId, chunk, context }) {
if (chunk.type === 'FINAL_RESPONSE') {
const quality = await assessQuality(chunk.finalResponseText);
if (quality.score < 0.5) {
return {
action: GuardrailAction.BLOCK,
reason: 'Response did not meet quality standards',
metadata: { qualityScore: quality.score, agent: sourceAgentId }
};
}
}
return null;
}
}
class SensitiveDataGuardrail implements ICrossAgentGuardrailService {
// Only observe agents handling sensitive data
observeAgentIds = ['data-analyst', 'report-generator'];
canInterruptOthers = true;
config = { evaluateStreamingChunks: true };
async evaluateCrossAgentOutput({ chunk }) {
if (chunk.textDelta && containsPII(chunk.textDelta)) {
return {
action: GuardrailAction.SANITIZE,
modifiedText: redactPII(chunk.textDelta)
};
}
return null;
}
}
Extends
Properties
canInterruptOthers?
optionalcanInterruptOthers:boolean
Defined in: packages/agentos/src/safety/guardrails/ICrossAgentGuardrailService.ts:158
Whether this guardrail can interrupt other agents' streams.
When true:
- GuardrailAction.BLOCK terminates the observed agent's stream
- GuardrailAction.SANITIZE modifies the observed agent's output
When false (default):
- Guardrail can only observe and log (FLAG action)
- BLOCK/SANITIZE actions are downgraded to FLAG
Default
false
config?
optionalconfig:GuardrailConfig
Defined in: packages/agentos/src/safety/guardrails/IGuardrailService.ts:521
Configuration for evaluation behavior. Controls streaming vs final-only evaluation and rate limiting.
Inherited from
id?
optionalid:string
Defined in: packages/agentos/src/safety/guardrails/IGuardrailService.ts:508
A stable id for this guard. A required guard is named by it (AgentOSConfig.requiredGuardrails), and it travels
with every verdict as metadata.guardrailId. Without one, a registered guard is known by its descriptor's id.
Inherited from
observeAgentIds?
optionalobserveAgentIds:string[]
Defined in: packages/agentos/src/safety/guardrails/ICrossAgentGuardrailService.ts:143
Agent IDs this guardrail observes.
- Empty array
[]or undefined: Observe all agents in the agency - Specific IDs: Only observe listed agents
Example
// Observe specific workers
observeAgentIds = ['worker-1', 'worker-2'];
// Observe all agents
observeAgentIds = [];
stages?
optionalstages: readonly ("output"|"input")[]
Defined in: packages/agentos/src/safety/guardrails/IGuardrailService.ts:515
The stages this guard evaluates, for a guard whose methods say more than its configuration: one that implements both methods but runs on one stage declares that stage. A required guard is held to what it declares, and a stage it does not declare is reported as missing instead of passing on the method's existence. Without it, a stage counts as covered when its method exists.
Inherited from
Methods
evaluateCrossAgentOutput()?
optionalevaluateCrossAgentOutput(payload):Promise<GuardrailEvaluationResult|null>
Defined in: packages/agentos/src/safety/guardrails/ICrossAgentGuardrailService.ts:175
Evaluate output from an observed agent.
Called when an observed agent (per observeAgentIds) emits a chunk. The evaluation timing depends on IGuardrailService.config:
evaluateStreamingChunks: true: Called for each TEXT_DELTAevaluateStreamingChunks: false: Called only for FINAL_RESPONSE
Parameters
payload
Cross-agent context and chunk to evaluate
Returns
Promise<GuardrailEvaluationResult | null>
Evaluation result, or null to allow without action
Remarks
- Only effective when canInterruptOthers is
true - Falls back to this guardrail's own stream for logging/metadata
evaluateInput()?
optionalevaluateInput(payload):Promise<GuardrailEvaluationResult|null>
Defined in: packages/agentos/src/safety/guardrails/IGuardrailService.ts:537
Evaluate user input before orchestration.
Called once per request before the orchestration pipeline starts. Use this to validate, sanitize, or block user messages.
Parameters
payload
Input and context to evaluate
Returns
Promise<GuardrailEvaluationResult | null>
Evaluation result, or null to allow without action
Remarks
- Return
BLOCKto prevent the request from being processed - Return
SANITIZEwithmodifiedTextto clean the input - Return
nullorALLOWto let the request through
Inherited from
IGuardrailService.evaluateInput
evaluateOutput()?
optionalevaluateOutput(payload):Promise<GuardrailEvaluationResult|null>
Defined in: packages/agentos/src/safety/guardrails/IGuardrailService.ts:562
Evaluate agent output before streaming to client.
Called for each chunk of a guarded output stream (the turn's, and the
continuation after an external tool result) that carries
isFinal: true (the FINAL_RESPONSE, an ERROR the turn yields), for each
TEXT_DELTA chunk when GuardrailConfig.evaluateStreamingChunks is
set (real-time filtering), and for each TOOL_CALL_REQUEST chunk when
GuardrailConfig.evaluateToolCalls is set. Other chunks are not
passed to it, nor is the ERROR that processRequest() yields from its own
catch.
Parameters
payload
Response chunk and context to evaluate
Returns
Promise<GuardrailEvaluationResult | null>
Evaluation result, or null to allow without action
Remarks
- Return
BLOCKto immediately terminate the stream with an error chunk, or with a FINAL_RESPONSE holdingreplacementTextwhen the result carries a non-empty one - Return
SANITIZEwithmodifiedTextto redact/modify content; on a TOOL_CALL_REQUEST chunk it is applied asBLOCK - Streaming evaluation adds latency; use only when real-time filtering is required