Agentic Patterns — Governance Placement

This page shows where TealTiger governance fits in common agentic architectures. Each pattern has specific governance insertion points and recommended controls.

Pattern 1: Single-Agent Tool Calling

Architecture: User → LLM → Tool Selection → Tool Execution → Response Governance insertion points:

Point	TealTiger Control	Why
Before tool execution	`evaluate()` with action + tool params	Enforce scope, cost, and safety
After LLM response	Content classification	Detect PII/secrets in output
Session boundary	TealTemporal session TTL	Prevent indefinite sessions

// Governance at tool boundary
const decision = await engine.evaluate(
  { action: 'tool.invoke', tool: toolName, toolParams: params },
  { correlation_id: uuid(), agentId: 'single-agent-001' }
);
if (decision.action !== 'ALLOW') return handleDenial(decision);
await executeTool(toolName, params);

Recommended controls: NHI scope enforcement, cost per-request limits, FREEZE rules for destructive tools.

Pattern 2: Multi-Agent Orchestration

Architecture: Orchestrator → Agent A → Agent B → Agent C → Merge → Response Governance insertion points:

Point	TealTiger Control	Why
Orchestrator dispatch	NHI scope check per agent	Each agent only does what it’s scoped for
Inter-agent communication	Content classification	Prevent injection propagation
Each agent’s tool calls	Per-agent `evaluate()`	Independent governance per agent
Merge/aggregation	Drift detection	Detect anomalous outputs

// Each agent has its own identity and scope
const agentADecision = await engine.evaluate(
  { action: 'research.search', tool: 'web_search' },
  { correlation_id, agentId: 'research-agent' }
);

const agentBDecision = await engine.evaluate(
  { action: 'code.write', tool: 'file_write', toolParams: { path: '/src/...' } },
  { correlation_id, agentId: 'coding-agent' }
);

Recommended controls: Per-agent NHI with separate scopes, TealFlow workflow orchestration, per-agent cost attribution, drift detection across agents.

Pattern 3: RAG (Retrieval-Augmented Generation)

Architecture: User → Query → Retriever → Context Assembly → LLM → Response Governance insertion points:

Point	TealTiger Control	Why
Retrieved content	Memory provenance tagging	Track trust level of each source
Context assembly	TealState context size governance	Prevent context overflow
Before LLM call	Content classification on assembled context	Detect injected instructions in retrieved docs
After LLM response	Output classification	Detect exfiltration attempts

// Tag retrieved content with provenance
const contextEntry = {
  content: retrievedDoc,
  source: 'web_search',
  trust_level: 'unverified',
};

// Evaluate context size before LLM call
const stateDecision = await engine.evaluate(
  { action: 'context.add', toolParams: { tokens: estimateTokens(contextEntry) } },
  { correlation_id, agentId: 'rag-agent' }
);

Recommended controls: TealState context limits, memory provenance (5-tier trust), markdown exfiltration detection, TealClassifier on retrieved content.

Pattern 4: AI Coding Agent

Architecture: User → Planning → Code Generation → File Write → Test → Commit Governance insertion points:

Point	TealTiger Control	Why
Planning phase	PLAN_ONLY mode	Evaluate without side effects
File write	CODE_CHANGE governance	Path/branch allowlists
Commit	Two-person rule (approval_required)	Human review for sensitive changes
All phases	FREEZE rules	Block writes to critical paths

// Planning phase — dry run
const plan = await engine.evaluate(
  { action: 'code.write', toolParams: { path: '/src/auth/login.ts' } },
  { correlation_id, agentId: 'coding-agent', mode: 'PLAN_ONLY' }
);

// Execution phase — full governance
const decision = await engine.evaluate(
  { action: 'CODE_CHANGE', toolParams: { path: '/src/auth/login.ts', diff_hash: sha256(diff) } },
  { correlation_id, agentId: 'coding-agent' }
);
// decision.action === 'PENDING' (approval_required for auth paths)

Recommended controls: CODE_CHANGE action class, path allowlists, branch allowlists, diff hash requirement, two-person rule, FREEZE on /etc/, .env, secrets/.

Pattern 5: Autonomous Background Agent

Architecture: Scheduled trigger → Agent → Multiple actions → Report Governance insertion points:

Point	TealTiger Control	Why
Agent startup	NHI attestation	Verify agent identity
Each action	Full evaluate() pipeline	Continuous governance
Session duration	TealTemporal session TTL	Prevent runaway agents
Time boundaries	Time-of-day restrictions	Block off-hours execution
Cost accumulation	Per-agent daily budget	Prevent cost explosion

// Background agent with temporal controls
const engine = new TealEngine({
  temporal: {
    session_ttl: '4h',  // Max 4 hours per run
    time_restrictions: [{
      actions: ['database.*'],
      allowed_windows: [{ days: ['mon','tue','wed','thu','fri'], start: '09:00', end: '17:00' }]
    }]
  },
  cost_governance: { daily_per_agent_max: 25.00 },
});

Recommended controls: Session TTL, time-of-day restrictions, cooldown periods, per-agent cost limits, drift detection (behavioral baseline).

Pattern 6: Human-in-the-Loop Agent

Architecture: User → Agent → Proposal → Human Review → Execute/Reject Governance insertion points:

Point	TealTiger Control	Why
Agent proposal	`automation: 'approval_required'`	Hold for human review
Human approval	`approveDecision(token)`	Cryptographic approval evidence
Execution	Full evaluate() after approval	Re-verify before execution

// Agent proposes, human approves
const decision = await engine.evaluate(
  { action: 'payment.refund', toolParams: { amount: 500, customer: 'cust-123' } },
  { correlation_id, agentId: 'support-bot' }
);

// decision.action === 'PENDING'
// Human reviews in dashboard...

// After approval:
const approved = await engine.approveDecision(decision.approval.token, {
  approvedBy: '[email protected]',
  justification: 'Verified refund request #RF-8821',
});

Recommended controls: Automation levels (approval_required for high-value actions), approval token TTL, TealProof receipts for non-repudiation.

Governance Placement Summary

Pattern	Primary Risk	Key TealTiger Controls
Single-agent	Ungoverned tool calls	NHI scope, FREEZE, cost limits
Multi-agent	Privilege escalation across agents	Per-agent NHI, TealFlow, drift
RAG	Context poisoning, exfiltration	TealState, provenance, classifier
Coding agent	Unauthorized code changes	CODE_CHANGE, PLAN_ONLY, two-person rule
Background agent	Runaway execution, cost explosion	Session TTL, time restrictions, cost ceiling
Human-in-the-loop	Approval bypass	approval_required, TealProof receipts

TealEngine v1.3

Full evaluation pipeline API

NHI Governance

Agent identity and scope enforcement

FREEZE Rules

Immutable safety controls

Orchestration Patterns

Detailed flow examples

Getting Started

Core Concepts

Governance Domains

Policy Authoring

Architecture

Integrations

Deployment

API Reference

Cookbook

Compliance & GRC

Guides

Reference

Versions

Playground

Dashboard

Blog

Agentic Patterns — Governance Placement

Pattern 1: Single-Agent Tool Calling

Pattern 2: Multi-Agent Orchestration

Pattern 3: RAG (Retrieval-Augmented Generation)

Pattern 4: AI Coding Agent

Pattern 5: Autonomous Background Agent

Pattern 6: Human-in-the-Loop Agent

Governance Placement Summary

TealEngine v1.3

NHI Governance

FREEZE Rules

Orchestration Patterns

​Pattern 1: Single-Agent Tool Calling

​Pattern 2: Multi-Agent Orchestration

​Pattern 3: RAG (Retrieval-Augmented Generation)

​Pattern 4: AI Coding Agent

​Pattern 5: Autonomous Background Agent

​Pattern 6: Human-in-the-Loop Agent

​Governance Placement Summary

​Related Documentation

TealEngine v1.3

NHI Governance

FREEZE Rules

Orchestration Patterns

Pattern 1: Single-Agent Tool Calling

Pattern 2: Multi-Agent Orchestration

Pattern 3: RAG (Retrieval-Augmented Generation)

Pattern 4: AI Coding Agent

Pattern 5: Autonomous Background Agent

Pattern 6: Human-in-the-Loop Agent

Governance Placement Summary

Related Documentation