Identity And Guardrails

Identity and Guardrails

Agent objective

Produce a builder with guardrails, behavioral contracts, and safety controls correctly configured so the agent operates within defined bounds and can be stopped when needed.

When to load this skill

• Deploying an agent in a multi-tenant or public-facing context
• Restricting which tools an agent can call or which topics it can address
• Requiring human approval before certain tool calls
• Needing runtime pause/resume/stop controls
• Enforcing token/iteration/output-length budgets at the contract level

Implementation baseline

import { ReactiveAgents } from "@reactive-agents/runtime";

const agent = await ReactiveAgents.create()
  .withName("assistant")
  .withProvider("anthropic")
  .withReasoning({ defaultStrategy: "adaptive" })
  .withTools({ allowedTools: ["web-search", "http-get", "file-read", "checkpoint"] })
  .withGuardrails({
    injection: true,     // detect prompt injection attempts
    pii: true,           // detect and mask PII in inputs/outputs
    toxicity: true,      // detect toxic content
  })
  .withBehavioralContracts({
    deniedTools: ["file-write", "shell-execute"],
    maxToolCalls: 30,
    maxIterations: 20,
    requireDisclosure: true,   // agent must disclose it is an AI
  })
  .withKillSwitch()   // enables pause/resume/stop/terminate controls
  .withIdentity()     // enables identity and session tracking
  .withAudit()        // records all tool calls and decisions to audit log
  .build();

Key patterns

Guardrails

.withGuardrails()
// Enables all detectors with defaults: injection=true, pii=true, toxicity=true

.withGuardrails({
  injection: true,           // detect "ignore previous instructions" attacks
  pii: false,                // disable PII masking (e.g., agent legitimately processes PII)
  toxicity: true,
  customBlocklist: ["competitor-product", "internal-codename"],  // substring blocklist
})

Guardrail violations abort the turn and return a structured error — the agent never processes the blocked content.

Behavioral contracts

Full field reference for .withBehavioralContracts(contract):

.withBehavioralContracts({
  deniedTools: ["file-delete", "shell-execute"],    // tools the agent may NEVER call
  allowedTools: ["web-search", "file-read"],        // if set, ONLY these tools are allowed
  maxToolCalls: 50,                                  // hard stop after N total tool calls
  maxIterations: 20,                                 // hard stop after N reasoning iterations
  maxOutputLength: 4000,                             // truncate/block output over N chars
  deniedTopics: ["competitor names", "legal advice"], // topics agent must refuse
  requireDisclosure: true,                           // first response must disclose AI identity
})

Contract violations are enforced at runtime — violations halt the current turn with a ContractViolation error.

Kill switch (runtime control)

.withKillSwitch()
// Enables runtime controls on the built agent handle:

const handle = agent.run("Do a long task...");

// Graceful pause (waits for current phase to finish)
await handle.pause();

// Resume from paused state
await handle.resume();

// Graceful stop (finishes current phase, then stops)
await handle.stop("User requested cancellation");

// Immediate termination
await handle.terminate("Emergency shutdown");

Kill switch controls are no-ops if .withKillSwitch() was not called during build.

Identity and audit

.withIdentity()   // enables agent identity headers, session IDs, and persona tracking
.withAudit()      // records all tool calls, guardrail decisions, and contract checks to an audit trail

Identity and audit work independently — enable both for full traceability.

GuardrailsOptions reference

BehavioralContract reference

Pitfalls

• .withGuardrails() with no args enables ALL detectors — disable selectively if your use case legitimately handles PII
• deniedTools in a contract and allowedTools in .withTools() are independent — a tool can be in .withTools({ allowedTools }) but still blocked by a contract's deniedTools
• Kill switch controls (pause, resume, stop, terminate) are no-ops without .withKillSwitch() — no error is thrown, calls are silently ignored
• requireDisclosure enforces the agent states it is AI in its first response — this is a prompt-level enforcement, not cryptographic
• Contract violations raise ContractViolation errors — handle these in your error callback or the agent run will throw
• .withAudit() without a log destination writes to the observability stream — add .withObservability() to capture audit events