Tutorial 20 — TypeScript SDK (@agentmesh/sdk)

Build governance-aware AI agents in TypeScript and Node.js. The @microsoft/agentmesh-sdk package provides cryptographic identity (Ed25519 DIDs), trust scoring, declarative policy evaluation, and hash-chain audit logging — all in a single npm install.

Prerequisites: Node.js ≥ 18 · TypeScript ≥ 5.4 Package: @microsoft/agentmesh-sdk v1.0.0 Modules: AgentIdentity, TrustManager, PolicyEngine, AuditLogger, AgentMeshClient

What you'll learn:

Section	Topic
Quick Start	Evaluate a policy in 5 lines of TypeScript
AgentMeshClient	Unified governance pipeline — identity + trust + policy + audit
PolicyEngine	Declarative rules, YAML policies, conflict resolution
AgentIdentity	Ed25519 key pairs, DIDs, delegation, capabilities
TrustManager	Bayesian trust scoring, tiers, decay
AuditLogger	Hash-chain audit logging and verification
Framework Integration	LangChain.js, OpenAI Node SDK
Configuration Reference	Defaults, environment variables, tuning
Error Handling	TypeScript-specific patterns
Cross-Reference	Equivalent Python tutorials

Installation

npm install @microsoft/agentmesh-sdk

The SDK has two runtime dependencies — @noble/ed25519 for cryptography and js-yaml for YAML policy parsing. Both are installed automatically.

For TypeScript projects, types are included — no separate @types/ package is needed.

# Verify the install
node -e "const sdk = require('@microsoft/agentmesh-sdk'); console.log(Object.keys(sdk))"

Quick Start

Five lines to evaluate your first policy:

import { PolicyEngine } from '@microsoft/agentmesh-sdk';

const engine = new PolicyEngine([
  { action: 'data.read',  effect: 'allow' },
  { action: 'data.write', effect: 'deny'  },
]);

console.log(engine.evaluate('data.read'));   // 'allow'
console.log(engine.evaluate('data.write'));  // 'deny'
console.log(engine.evaluate('data.delete')); // 'deny'  ← default when no rule matches

Tip: The default decision when no rule matches is 'deny' — secure by default.

Or use the unified AgentMeshClient for the full governance pipeline:

import { AgentMeshClient } from '@microsoft/agentmesh-sdk';

const client = AgentMeshClient.create('my-agent', {
  capabilities: ['data.read', 'data.write'],
  policyRules: [
    { action: 'data.read',  effect: 'allow' },
    { action: 'data.write', effect: 'allow', conditions: { role: 'admin' } },
    { action: '*',           effect: 'deny'  },
  ],
});

const result = await client.executeWithGovernance('data.read');
console.log(result.decision);    // 'allow'
console.log(result.trustScore);  // { overall: 0.5, tier: 'Provisional', ... }
console.log(result.auditEntry);  // { hash: '3a7f...', previousHash: '0000...', ... }

AgentMeshClient

The AgentMeshClient is the recommended entry point. It wires together identity, trust, policy, and audit into a single governance-aware pipeline.

Creating a Client

import { AgentMeshClient } from '@microsoft/agentmesh-sdk';

// Quick creation with defaults
const client = AgentMeshClient.create('sales-agent', {
  capabilities: ['crm.read', 'crm.write', 'email.send'],
});

// Access individual components
console.log(client.identity.did);    // did:agentmesh:sales-agent:<fingerprint>
console.log(client.trust);           // TrustManager instance
console.log(client.policy);          // PolicyEngine instance
console.log(client.audit);           // AuditLogger instance

Full Configuration

import { AgentMeshClient, AgentMeshConfig } from '@microsoft/agentmesh-sdk';

const config: AgentMeshConfig = {
  agentId: 'analytics-agent',
  capabilities: ['data.read', 'report.generate'],
  trust: {
    initialScore: 0.6,
    decayFactor: 0.98,
    thresholds: {
      untrusted: 0.0,
      provisional: 0.3,
      trusted: 0.6,
      verified: 0.85,
    },
  },
  policyRules: [
    { action: 'data.read',        effect: 'allow' },
    { action: 'report.generate',  effect: 'allow' },
    { action: '*',                 effect: 'deny'  },
  ],
  audit: {
    maxEntries: 50_000,
  },
};

const client = new AgentMeshClient(config);

The Governance Pipeline

executeWithGovernance() runs every action through a four-stage pipeline:

  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐    ┌──────────────┐
  │ 1. Policy    │───▶│ 2. Trust     │───▶│ 3. Audit     │───▶│ 4. Trust     │
  │    Evaluate  │    │    Score     │    │    Log       │    │    Update    │
  └──────────────┘    └──────────────┘    └──────────────┘    └──────────────┘

const result = await client.executeWithGovernance('data.read', {
  userId: 'alice',
  department: 'engineering',
});

// GovernanceResult
console.log(result.decision);       // 'allow' | 'deny' | 'review'
console.log(result.trustScore);     // TrustScore object
console.log(result.auditEntry);     // AuditEntry with hash chain
console.log(result.executionTime);  // milliseconds (e.g., 1.234)

GovernanceResult shape:

Field	Type	Description
`decision`	`'allow' \| 'deny' \| 'review'`	Policy evaluation outcome
`trustScore`	`TrustScore`	Current trust score after action
`auditEntry`	`AuditEntry`	Immutable audit record created
`executionTime`	`number`	Pipeline duration in milliseconds

PolicyEngine

The PolicyEngine supports two modes: legacy flat rules (simple action matching) and rich YAML/JSON policies (expressions, rate limits, approvals, conflict resolution).

§4.1 Legacy Flat Rules

Flat rules match an action string and return an effect:

import { PolicyEngine, PolicyRule } from '@microsoft/agentmesh-sdk';

const rules: PolicyRule[] = [
  { action: 'data.read',   effect: 'allow' },
  { action: 'data.write',  effect: 'allow', conditions: { role: 'admin' } },
  { action: 'data.*',      effect: 'deny'  },
  { action: '*',            effect: 'deny'  },
];

const engine = new PolicyEngine(rules);

engine.evaluate('data.read');                      // 'allow'
engine.evaluate('data.write', { role: 'admin' });  // 'allow'
engine.evaluate('data.write', { role: 'viewer' }); // 'deny' — conditions don't match
engine.evaluate('data.delete');                    // 'deny' — wildcard catch-all

Evaluation order: First matching rule wins. Place specific rules before wildcards.

§4.2 Loading Rules from YAML

const engine = new PolicyEngine();

// Load from a YAML file on disk
await engine.loadFromYAML('./policies/production.yaml');

engine.evaluate('data.read');  // Uses loaded rules

§4.3 Rich Policy Documents

Rich policies add expressions, rate limits, approval workflows, and conflict resolution:

import { PolicyEngine } from '@microsoft/agentmesh-sdk';

const engine = new PolicyEngine();

const policy = engine.loadYaml(`
apiVersion: governance.toolkit/v1
name: data-access-policy
description: Controls data operations for analytics agents
agents:
  - 'did:agentmesh:analytics-*'
scope: tenant

rules:
  - name: admin-full-access
    condition: "user.role == 'admin'"
    ruleAction: allow
    priority: 100

  - name: analyst-read-only
    condition: "user.role in ['analyst', 'viewer']"
    ruleAction: allow
    priority: 50

  - name: rate-limit-writes
    condition: "action == 'data.write'"
    ruleAction: allow
    priority: 75
    limit: '100/hour'

  - name: require-approval-for-delete
    condition: "action == 'data.delete'"
    ruleAction: require_approval
    priority: 90
    approvers:
      - 'did:agentmesh:security-team'
      - 'did:agentmesh:data-owner'

  - name: default-deny
    ruleAction: deny
    priority: 0

default_action: deny
`);

console.log(policy.name);  // 'data-access-policy'

§4.4 Rich Policy Evaluation

const result = engine.evaluatePolicy(
  'did:agentmesh:analytics-agent:abc123',
  { user: { role: 'admin' }, action: 'data.write' }
);

console.log(result.allowed);       // true
console.log(result.action);        // 'allow'
console.log(result.matchedRule);   // 'admin-full-access'
console.log(result.policyName);    // 'data-access-policy'
console.log(result.rateLimited);   // false
console.log(result.approvers);     // []
console.log(result.evaluationMs);  // 0.123

PolicyDecisionResult shape:

Field	Type	Description
`allowed`	`boolean`	Whether the action is permitted
`action`	`PolicyAction`	`'allow'` \| `'deny'` \| `'warn'` \| `'require_approval'` \| `'log'`
`matchedRule`	`string?`	Name of the rule that matched
`policyName`	`string?`	Name of the policy that matched
`reason`	`string?`	Human-readable explanation
`approvers`	`string[]`	Required approver DIDs (for `require_approval`)
`rateLimited`	`boolean`	Whether rate limiting is active
`evaluatedAt`	`Date`	Timestamp of evaluation
`evaluationMs`	`number?`	Duration of evaluation in ms

§4.5 Expression Syntax

Rich policy conditions use a simple expression language:

Operator	Example	Description
`==`	`user.role == 'admin'`	Equality (string, number, boolean)
`!=`	`status != 'blocked'`	Not equal
`>`	`token_count > 2048`	Greater than
`<`	`risk_score < 0.5`	Less than
`>=`	`trust_level >= 0.85`	Greater than or equal
`<=`	`attempts <= 3`	Less than or equal
`in`	`role in ['admin', 'analyst']`	Membership in list
`not in`	`env not in ['prod']`	Not a member
`and`	`role == 'admin' and dept == 'eng'`	Logical AND
`or`	`level > 5 or isVerified`	Logical OR
(truthy)	`isVerified`	Property existence / truthiness

Nested path access: Use dots to access nested properties — user.role, request.headers.authorization, agent.trust.tier.

§4.6 Rate Limiting

Add a limit field to any rule to enforce rate limits:

rules:
  - name: throttle-api-calls
    condition: "action == 'api.call'"
    ruleAction: allow
    limit: '1000/minute'

Supported time windows:

Format	Example
`N/second`	`100/second`
`N/minute`	`1000/minute`
`N/hour`	`10000/hour`
`N/day`	`100000/day`

§4.7 Conflict Resolution

When multiple policies apply to the same action, the PolicyConflictResolver determines which rule wins:

import {
  PolicyEngine,
  PolicyConflictResolver,
  ConflictResolutionStrategy,
} from '@microsoft/agentmesh-sdk';

const engine = new PolicyEngine([], ConflictResolutionStrategy.DenyOverrides);

Strategies:

Strategy	Enum Value	Behaviour
Deny Overrides	`deny_overrides`	Any `deny` rule wins — safety first
Allow Overrides	`allow_overrides`	Any `allow` rule wins — permissive
Priority First Match	`priority_first_match`	Highest `priority` value wins (default)
Most Specific Wins	`most_specific_wins`	Most specific `scope` + highest `priority` wins

const resolver = new PolicyConflictResolver(
  ConflictResolutionStrategy.MostSpecificWins
);

const result = resolver.resolve([
  {
    action: 'allow',
    priority: 50,
    scope: PolicyScope.Global,
    policyName: 'global-policy',
    ruleName: 'allow-reads',
    reason: 'Global read access',
    approvers: [],
  },
  {
    action: 'deny',
    priority: 50,
    scope: PolicyScope.Agent,
    policyName: 'agent-policy',
    ruleName: 'deny-untrusted',
    reason: 'Agent-specific restriction',
    approvers: [],
  },
]);

console.log(result.winningDecision.action);   // 'deny' — agent scope is more specific
console.log(result.strategyUsed);             // 'most_specific_wins'
console.log(result.conflictDetected);         // true
console.log(result.resolutionTrace);          // step-by-step resolution log

Scope specificity order (most → least specific):

  Agent  >  Tenant  >  Global

§4.8 Managing Multiple Policies

const engine = new PolicyEngine();

// Load multiple policies
engine.loadYaml(securityPolicyYaml);
engine.loadYaml(compliancePolicyYaml);
engine.loadYaml(operationalPolicyYaml);

// List and inspect
console.log(engine.listPolicies());        // ['security', 'compliance', 'operational']
console.log(engine.getPolicy('security')); // Policy object

// Remove a policy
engine.removePolicy('operational');

// Clear all policies
engine.clearPolicies();

AgentIdentity

Each agent gets a cryptographic identity backed by Ed25519 key pairs. The identity produces a DID (Decentralized Identifier) and supports signing, verification, delegation, and lifecycle management.

§5.1 Generating an Identity

import { AgentIdentity } from '@microsoft/agentmesh-sdk';

const agent = AgentIdentity.generate('sales-assistant', ['crm.read', 'email.send'], {
  name: 'Sales Assistant',
  description: 'Handles inbound sales inquiries',
  organization: 'Contoso',
  sponsor: 'alice@contoso.com',
  expiresAt: new Date('2026-01-01'),
});

console.log(agent.did);           // did:agentmesh:sales-assistant:<fingerprint>
console.log(agent.publicKey);     // Uint8Array (Ed25519 DER)
console.log(agent.capabilities);  // ['crm.read', 'email.send']
console.log(agent.status);        // 'active'
console.log(agent.organization);  // 'Contoso'
console.log(agent.sponsor);       // 'alice@contoso.com'

DID format: did:agentmesh:<agentId>:<fingerprint>

The fingerprint is derived from the public key, making each DID globally unique and cryptographically verifiable.

§5.2 Signing and Verification

const message = new TextEncoder().encode('Transfer $500 to account 1234');

// Sign with private key
const signature = agent.sign(message);

// Verify with public key
const valid = agent.verify(message, signature);
console.log(valid);  // true

// Tampered data fails verification
const tampered = new TextEncoder().encode('Transfer $50000 to account 9999');
console.log(agent.verify(tampered, signature));  // false

Note: Signing requires the private key. Identities imported via fromJSON() without a privateKey field can only verify, not sign.

§5.3 Capability Checking

Capabilities support exact matching and wildcard patterns:

const agent = AgentIdentity.generate('worker', [
  'data.read',
  'data.write',
  'report.generate',
]);

agent.hasCapability('data.read');         // true
agent.hasCapability('data.delete');       // false
agent.hasCapability('data.*');            // true  — wildcard matches data.read & data.write
agent.hasCapability('report.generate');   // true

§5.4 Delegation

Create child identities with narrowed capabilities:

const parent = AgentIdentity.generate('orchestrator', [
  'data.read', 'data.write', 'admin', 'deploy',
]);

// Child can only read and write — no admin or deploy
const child = parent.delegate('data-worker', ['data.read', 'data.write'], {
  description: 'Scoped worker for data pipeline',
  sponsor: 'pipeline-team@contoso.com',
});

console.log(child.parentDid);         // parent's DID
console.log(child.delegationDepth);   // 1
console.log(child.hasCapability('data.read'));   // true
console.log(child.hasCapability('admin'));       // false — not delegated

// Delegation chains
const grandchild = child.delegate('read-only', ['data.read']);
console.log(grandchild.delegationDepth);  // 2

§5.5 Lifecycle Management

const agent = AgentIdentity.generate('temp-agent', ['task.run']);

console.log(agent.isActive());  // true
console.log(agent.status);      // 'active'

// Suspend — temporary, reversible
agent.suspend('Under investigation');
console.log(agent.isActive());  // false
console.log(agent.status);      // 'suspended'

// Reactivate
agent.reactivate();
console.log(agent.isActive());  // true

// Revoke — permanent, irreversible
agent.revoke('Compromised credentials');
console.log(agent.status);      // 'revoked'

// Cannot reactivate a revoked identity
try {
  agent.reactivate();  // throws
} catch (e) {
  console.log(e.message);  // Cannot reactivate a revoked identity
}

State machine:

  generate()      suspend()           revoke()
      │               │                  │
      ▼               ▼                  ▼
  ┌────────┐     ┌───────────┐     ┌─────────┐
  │ active │────▶│ suspended │     │ revoked │
  └────────┘     └───────────┘     └─────────┘
       │          reactivate() │         ▲
       │              │                  │
       │              ▼                  │
       │          ┌────────┐             │
       └─────────▶│ active │─────────────┘
                  └────────┘  revoke()

§5.6 Serialization

JSON round-trip:

// Export
const json = agent.toJSON();
// {
//   did: 'did:agentmesh:sales-assistant:...',
//   publicKey: 'base64...',
//   privateKey: 'base64...',
//   capabilities: ['crm.read', 'email.send'],
//   name: 'Sales Assistant',
//   status: 'active',
//   createdAt: '2025-01-15T...',
//   ...
// }

// Import
const restored = AgentIdentity.fromJSON(json);
console.log(restored.did === agent.did);  // true

JWK / JWKS (RFC 7517):

// Export as JWK (public key only)
const jwk = agent.toJWK(false);
// { kty: 'OKP', crv: 'Ed25519', x: '...', kid: 'did:agentmesh:...' }

// Export with private key
const jwkPrivate = agent.toJWK(true);

// JWK Set
const jwks = agent.toJWKS(false);
// { keys: [{ kty: 'OKP', crv: 'Ed25519', ... }] }

// Import from JWK
const fromJwk = AgentIdentity.fromJWK(jwk);

// Import from JWKS (picks first key, or by kid)
const fromJwks = AgentIdentity.fromJWKS(jwks, agent.did);

W3C DID Document:

const didDoc = agent.toDIDDocument();
// {
//   '@context': 'https://www.w3.org/ns/did/v1',
//   id: 'did:agentmesh:sales-assistant:...',
//   verificationMethod: [...],
//   authentication: [...],
//   ...
// }

§5.7 Identity Registry

Manage multiple identities with the IdentityRegistry:

import { AgentIdentity, IdentityRegistry } from '@microsoft/agentmesh-sdk';

const registry = new IdentityRegistry();

const agent1 = AgentIdentity.generate('agent-1', ['read'], {
  sponsor: 'alice@contoso.com',
});
const agent2 = AgentIdentity.generate('agent-2', ['write'], {
  sponsor: 'alice@contoso.com',
});
const agent3 = AgentIdentity.generate('agent-3', ['admin'], {
  sponsor: 'bob@contoso.com',
});

registry.register(agent1);
registry.register(agent2);
registry.register(agent3);

console.log(registry.size);  // 3

// Look up by DID
const found = registry.get(agent1.did);
console.log(found?.did);  // agent1's DID

// Find all identities by sponsor
const aliceAgents = registry.getBySponsor('alice@contoso.com');
console.log(aliceAgents.length);  // 2

// List all active identities
console.log(registry.listActive().length);  // 3

// Revoke — cascades to all delegates
registry.revoke(agent1.did, 'Decommissioned');
console.log(registry.listActive().length);  // 2

TrustManager

The TrustManager maintains a Bayesian-inspired trust score for each peer agent. Scores decay over time and update based on interaction outcomes.

§6.1 Creating a Trust Manager

import { TrustManager, TrustConfig } from '@microsoft/agentmesh-sdk';

// Default configuration
const tm = new TrustManager();

// Custom configuration
const custom = new TrustManager({
  initialScore: 0.6,
  decayFactor: 0.98,
  thresholds: {
    untrusted: 0.0,
    provisional: 0.3,
    trusted: 0.6,
    verified: 0.85,
  },
});

Default values:

Property	Default	Description
`initialScore`	`0.5`	Starting score for unknown agents
`decayFactor`	`0.95`	Hourly decay multiplier
`thresholds.untrusted`	`0.0`	Floor for untrusted tier
`thresholds.provisional`	`0.3`	Entry point for provisional trust
`thresholds.trusted`	`0.6`	Entry point for trusted tier
`thresholds.verified`	`0.85`	Entry point for verified tier

§6.2 Recording Interactions

const tm = new TrustManager();

// Record successful interactions (default reward: 0.05)
tm.recordSuccess('peer-agent-1');
tm.recordSuccess('peer-agent-1');
tm.recordSuccess('peer-agent-1');

// Record failure (default penalty: 0.1)
tm.recordFailure('peer-agent-1');

// Custom reward / penalty values
tm.recordSuccess('peer-agent-2', 0.1);   // larger reward
tm.recordFailure('peer-agent-2', 0.2);   // harsher penalty

§6.3 Reading Trust Scores

const score = tm.getTrustScore('peer-agent-1');

console.log(score.overall);                // 0.55  (0-1, rounded to 3 decimals)
console.log(score.tier);                   // 'Provisional'
console.log(score.dimensions.reliability); // 0.75  (3 successes / 4 total)
console.log(score.dimensions.consistency); // 0.55  (current score after updates)

TrustScore shape:

interface TrustScore {
  overall: number;       // 0-1, rounded to 3 decimals
  dimensions: {
    reliability: number; // success ratio: successes / total interactions
    consistency: number; // current running score
  };
  tier: TrustTier;       // 'Untrusted' | 'Provisional' | 'Trusted' | 'Verified'
}

§6.4 Trust Tiers

Tiers map numeric scores to access levels:

Tier	Score Range	Typical Access
Untrusted	0.00 – 0.29	Blocked or heavily restricted
Provisional	0.30 – 0.59	Limited access, monitored
Trusted	0.60 – 0.84	Standard access
Verified	0.85 – 1.00	Full access, elevated privileges

Untrusted      ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  0.00 – 0.29
Provisional    ████████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  0.30 – 0.59
Trusted        ████████████████████████░░░░░░░░░░░░░░░░░░  0.60 – 0.84
Verified       ████████████████████████████████████████░░  0.85 – 1.00

§6.5 Peer Verification

Verify another agent's identity and establish initial trust:

import { AgentIdentity, TrustManager } from '@microsoft/agentmesh-sdk';

const tm = new TrustManager();
const peer = AgentIdentity.generate('remote-agent', ['data.read']);

const result = await tm.verifyPeer('remote-agent', peer);

console.log(result.verified);    // true — signature self-consistency checks passed
console.log(result.trustScore);  // initial TrustScore
console.log(result.reason);      // undefined when successful

§6.6 Trust Decay

Trust scores decay over time when there are no interactions. The decayFactor is applied hourly:

// With decayFactor: 0.95
// After 1 hour of inactivity: score × 0.95
// After 2 hours: score × 0.95²
// After 24 hours: score × 0.95²⁴ ≈ score × 0.292

const tm = new TrustManager({ decayFactor: 0.95 });
tm.recordSuccess('agent-x');
tm.recordSuccess('agent-x');

// Score starts high, but decays if no further interactions
const score = tm.getTrustScore('agent-x');
console.log(score.overall);  // Score reflects elapsed time since last interaction

Tip: Set decayFactor closer to 1.0 (e.g., 0.99) for slower decay, or lower (e.g., 0.90) for aggressive decay requiring frequent re-validation.

AuditLogger

The AuditLogger provides an append-only, hash-chain-linked audit trail. Each entry's SHA-256 hash incorporates the previous entry's hash, creating a tamper-evident chain similar to a blockchain.

§7.1 Logging Events

import { AuditLogger } from '@microsoft/agentmesh-sdk';

const logger = new AuditLogger();

const entry = logger.log({
  agentId: 'sales-assistant',
  action: 'crm.read',
  decision: 'allow',
});

console.log(entry.timestamp);     // ISO 8601 string
console.log(entry.hash);          // SHA-256 of this entry
console.log(entry.previousHash);  // SHA-256 of previous entry (genesis: '0' × 64)

§7.2 Hash-Chain Integrity

const logger = new AuditLogger();

logger.log({ agentId: 'agent-1', action: 'data.read',   decision: 'allow' });
logger.log({ agentId: 'agent-1', action: 'data.write',  decision: 'deny'  });
logger.log({ agentId: 'agent-2', action: 'report.send', decision: 'allow' });

// Verify the entire chain
console.log(logger.verify());  // true — all hashes are consistent

console.log(logger.length);   // 3

How the chain works:

  ┌─────────────────────────────────────────────────────────────────────────┐
  │  Entry 0            Entry 1            Entry 2                        │
  │  ┌──────────┐       ┌──────────┐       ┌──────────┐                   │
  │  │ hash: A  │──────▶│ prev: A  │──────▶│ prev: B  │                   │
  │  │ prev: 0⁶⁴│       │ hash: B  │       │ hash: C  │                   │
  │  └──────────┘       └──────────┘       └──────────┘                   │
  │                                                                       │
  │  Genesis hash = '000...000' (64 zeros)                                │
  │  Each hash = SHA-256(timestamp + agentId + action + decision + prev)  │
  └─────────────────────────────────────────────────────────────────────────┘

If any entry is tampered with, verify() returns false because the hash chain breaks.

§7.3 Querying the Log

// All entries for a specific agent
const agentEntries = logger.getEntries({ agentId: 'agent-1' });
console.log(agentEntries.length);  // 2

// Filter by action
const reads = logger.getEntries({ action: 'data.read' });

// Filter by time
const recent = logger.getEntries({
  since: new Date('2025-01-15T00:00:00Z'),
});

// Combine filters
const agentReads = logger.getEntries({
  agentId: 'agent-1',
  action: 'data.read',
  since: new Date('2025-01-01'),
});

§7.4 Exporting the Audit Trail

// Export as JSON string
const json = logger.exportJSON();
console.log(json);
// [
//   {
//     "timestamp": "2025-01-15T10:30:00.000Z",
//     "agentId": "agent-1",
//     "action": "data.read",
//     "decision": "allow",
//     "hash": "3a7f...",
//     "previousHash": "0000...0000"
//   },
//   ...
// ]

§7.5 Configuration

// Limit log size (oldest entries evicted when limit reached)
const logger = new AuditLogger({ maxEntries: 50_000 });

Option	Default	Description
`maxEntries`	`10,000`	Maximum entries before eviction

Framework Integration

§8.1 LangChain.js

Wrap LangChain tool calls with governance checks:

import { AgentMeshClient } from '@microsoft/agentmesh-sdk';
import { ChatOpenAI } from '@langchain/openai';
import { DynamicTool } from '@langchain/core/tools';

const client = AgentMeshClient.create('langchain-agent', {
  capabilities: ['search', 'calculate'],
  policyRules: [
    { action: 'search',    effect: 'allow' },
    { action: 'calculate', effect: 'allow' },
    { action: '*',          effect: 'deny'  },
  ],
});

// Governance-aware tool wrapper
function governedTool(name: string, fn: (input: string) => Promise<string>) {
  return new DynamicTool({
    name,
    description: `Governed: ${name}`,
    func: async (input: string) => {
      const gov = await client.executeWithGovernance(name, { input });

      if (gov.decision !== 'allow') {
        return `Action '${name}' denied by policy: ${gov.auditEntry.action}`;
      }
      return fn(input);
    },
  });
}

const searchTool = governedTool('search', async (query) => {
  // your search implementation
  return `Results for: ${query}`;
});

§8.2 OpenAI Node SDK

Add governance to OpenAI function calls:

import { AgentMeshClient } from '@microsoft/agentmesh-sdk';
import OpenAI from 'openai';

const client = AgentMeshClient.create('openai-agent', {
  capabilities: ['chat', 'function_call'],
  policyRules: [
    { action: 'chat',          effect: 'allow' },
    { action: 'function_call', effect: 'allow', conditions: { trusted: true } },
    { action: '*',              effect: 'deny'  },
  ],
});

const openai = new OpenAI();

async function governedCompletion(
  messages: OpenAI.ChatCompletionMessageParam[],
  tools?: OpenAI.ChatCompletionTool[]
) {
  // Check policy before calling OpenAI
  const gov = await client.executeWithGovernance('chat', {
    messageCount: messages.length,
    hasTools: !!tools,
  });

  if (gov.decision !== 'allow') {
    throw new Error(`Governance denied: ${gov.decision}`);
  }

  const response = await openai.chat.completions.create({
    model: 'gpt-4o',
    messages,
    tools,
  });

  // If the model wants to call a function, check that too
  const toolCalls = response.choices[0]?.message?.tool_calls;
  if (toolCalls) {
    for (const call of toolCalls) {
      const fnGov = await client.executeWithGovernance('function_call', {
        functionName: call.function.name,
        trusted: true,
      });

      if (fnGov.decision !== 'allow') {
        console.warn(`Function call '${call.function.name}' denied`);
      }
    }
  }

  return response;
}

Configuration Reference

§9.1 Default Values

Component	Property	Default
`PolicyEngine`	Conflict strategy	`PriorityFirstMatch`
`PolicyEngine`	Default action	`'deny'`
`PolicyRule`	`priority`	`0`
`PolicyRule`	`enabled`	`true`
`TrustManager`	`initialScore`	`0.5`
`TrustManager`	`decayFactor`	`0.95`
`TrustManager`	Success reward	`0.05`
`TrustManager`	Failure penalty	`0.1`
`AuditLogger`	`maxEntries`	`10,000`
`AuditLogger`	Genesis hash	`'0' × 64`
`AgentIdentity`	`status`	`'active'`
`AgentIdentity`	`delegationDepth`	`0` (root)

§9.2 Environment Variables

The SDK itself is configuration-object driven, but you can wire environment variables into your configuration:

import { AgentMeshClient } from '@microsoft/agentmesh-sdk';

const client = AgentMeshClient.create(
  process.env.AGENT_ID ?? 'default-agent',
  {
    capabilities: (process.env.AGENT_CAPABILITIES ?? 'read').split(','),
    trust: {
      initialScore: parseFloat(process.env.TRUST_INITIAL_SCORE ?? '0.5'),
      decayFactor: parseFloat(process.env.TRUST_DECAY_FACTOR ?? '0.95'),
    },
    audit: {
      maxEntries: parseInt(process.env.AUDIT_MAX_ENTRIES ?? '10000', 10),
    },
  }
);

Recommended environment variables:

Variable	Example	Maps to
`AGENT_ID`	`sales-assistant`	`agentId`
`AGENT_CAPABILITIES`	`read,write,search`	`capabilities` (comma-separated)
`TRUST_INITIAL_SCORE`	`0.6`	`trust.initialScore`
`TRUST_DECAY_FACTOR`	`0.98`	`trust.decayFactor`
`AUDIT_MAX_ENTRIES`	`50000`	`audit.maxEntries`
`POLICY_YAML_PATH`	`./policies/prod.yaml`	Path for `loadFromYAML()`

§9.3 TypeScript Configuration

The SDK targets ES2020 and uses Node16 module resolution. Ensure your tsconfig.json includes:

{
  "compilerOptions": {
    "target": "ES2020",
    "module": "Node16",
    "moduleResolution": "Node16",
    "strict": true,
    "esModuleInterop": true,
    "resolveJsonModule": true
  }
}

Error Handling

§10.1 Identity Errors

import { AgentIdentity } from '@microsoft/agentmesh-sdk';

// Revoked identity cannot be reactivated
const agent = AgentIdentity.generate('temp', ['read']);
agent.revoke('Done');

try {
  agent.reactivate();
} catch (error) {
  console.error(error.message);  // Cannot reactivate a revoked identity
}

// Duplicate registration
const registry = new IdentityRegistry();
registry.register(agent);

try {
  registry.register(agent);  // same DID
} catch (error) {
  console.error(error.message);  // Identity already registered
}

§10.2 Policy Errors

import { PolicyEngine } from '@microsoft/agentmesh-sdk';

const engine = new PolicyEngine();

// Invalid YAML
try {
  engine.loadYaml('not: valid: yaml: {{');
} catch (error) {
  console.error('YAML parse error:', error.message);
}

// Missing policy file
try {
  await engine.loadFromYAML('./nonexistent.yaml');
} catch (error) {
  console.error('File not found:', error.message);
}

§10.3 Governance Pipeline Errors

import { AgentMeshClient } from '@microsoft/agentmesh-sdk';

const client = AgentMeshClient.create('agent', {
  policyRules: [{ action: '*', effect: 'deny' }],
});

// Wrap executeWithGovernance in try/catch
try {
  const result = await client.executeWithGovernance('risky.action');

  if (result.decision === 'deny') {
    console.log('Action denied — audit trail:', result.auditEntry.hash);
  } else if (result.decision === 'review') {
    console.log('Action requires review — trust:', result.trustScore.tier);
  }
} catch (error) {
  console.error('Governance pipeline error:', error);
}

§10.4 TypeScript-Specific Patterns

Use type narrowing and discriminated unions for safe handling:

import type {
  PolicyDecisionResult,
  GovernanceResult,
  TrustTier,
} from '@microsoft/agentmesh-sdk';

// Type guard for trust tiers
function requiresTrust(tier: TrustTier, minimum: TrustTier): boolean {
  const order: TrustTier[] = ['Untrusted', 'Provisional', 'Trusted', 'Verified'];
  return order.indexOf(tier) >= order.indexOf(minimum);
}

// Use with governance results
async function safeExecute(
  client: AgentMeshClient,
  action: string
): Promise<GovernanceResult> {
  const result = await client.executeWithGovernance(action);

  if (result.decision === 'deny') {
    throw new Error(`Denied: ${action} (audit: ${result.auditEntry.hash})`);
  }

  if (!requiresTrust(result.trustScore.tier, 'Trusted')) {
    console.warn(`Low trust for ${action}: ${result.trustScore.tier}`);
  }

  return result;
}

Cross-Reference

The TypeScript SDK mirrors the Python packages. Use this table to find the equivalent Python tutorial for each topic:

TypeScript SDK	Python Package	Tutorial
`PolicyEngine`	`agent_os.policies`	Tutorial 01 — Policy Engine
`AgentIdentity`, `IdentityRegistry`	`agent_os.identity`	Tutorial 02 — Trust & Identity
`TrustManager`	`agent_os.trust`	Tutorial 02 — Trust & Identity
Framework integration	`agent_os.integrations`	Tutorial 03 — Framework Integrations
`AuditLogger`	`agent_os.audit`	Tutorial 04 — Audit & Compliance

Note: The TypeScript SDK wraps all governance features into a single @microsoft/agentmesh-sdk package, while the Python implementation splits them across separate agent_os.* modules. The APIs are designed for cross-language parity — policy YAML files work identically in both.

Source Files

Component	Location
Main exports	`packages/agent-mesh/sdks/typescript/src/index.ts`
Type definitions	`packages/agent-mesh/sdks/typescript/src/types.ts`
`AgentIdentity`	`packages/agent-mesh/sdks/typescript/src/identity.ts`
`TrustManager`	`packages/agent-mesh/sdks/typescript/src/trust.ts`
`PolicyEngine`	`packages/agent-mesh/sdks/typescript/src/policy.ts`
`AuditLogger`	`packages/agent-mesh/sdks/typescript/src/audit.ts`
`AgentMeshClient`	`packages/agent-mesh/sdks/typescript/src/client.ts`
Tests	`packages/agent-mesh/sdks/typescript/tests/*.test.ts`
Package config	`packages/agent-mesh/sdks/typescript/package.json`

Next Steps

Run the tests to see the SDK in action:

cd packages/agent-mesh/sdks/typescript
npm install && npm test

Load a YAML policy from the repository's policies/ directory and evaluate it against your agent
Build a multi-agent system using IdentityRegistry + TrustManager to establish peer trust between cooperating agents
Add audit export to your CI/CD pipeline — call logger.exportJSON() and persist the trail for compliance reviews
Explore the Python SDK tutorials (01–04) for equivalent patterns in Python

FilesExpand file tree

20-typescript-sdk.md

Latest commit

History