QWED-Agent Specification v1.0 - QWED Documentation

Status: Draft
Version: 1.1.0
Date: 2026-04-03
Extends: QWED-SPEC v1.0, QWED-Attestation v1.0

Introduction
Agent Verification Model
Agent Registration
Verification Requests
Tool Verification
Budget & Limits
Audit Trail
Trust Levels
Runtime Hardening
Implementation Guidelines

1. Introduction

1.1 Purpose

QWED-Agent defines a protocol for AI agents to verify their actions before execution. As agentic AI systems become more autonomous, QWED-Agent provides guardrails ensuring agents operate within defined boundaries.

1.2 Problem Statement

Problem	Risk
Agents execute unverified code	Security vulnerabilities
Agents make unverified calculations	Financial errors
Agents generate unverified SQL	Data corruption
Agents exceed resource limits	Cost overruns
No audit trail of agent actions	Compliance violations

1.3 Solution

QWED-Agent establishes:

Pre-execution verification of agent outputs
Tool call approval workflow
Budget enforcement
Complete audit trail
Trust level management

1.4 Terminology

Term	Definition
Agent	Autonomous AI system performing tasks
Principal	Entity that owns/controls the agent
Tool	External capability an agent can invoke
Action	Any operation an agent wants to perform
Verification Gate	Check before action execution
Budget	Resource limits for the agent

2. Agent Verification Model

2.1 Verification Flow

2.2 Verification Types for Agents

Action Type	Verification Engine	Risk Level
Math calculation	Math Engine	Low
Database query	SQL Engine	High
Code execution	Code Engine	Critical
External API call	Tool Verification	Medium
File operations	Security Check	High
Network requests	Policy Check	Medium

2.3 Decision Matrix

Verification	Risk level	Action
VERIFIED	Low	Execute immediately
VERIFIED	High	Execute with attestation
FAILED	Any	Block and notify principal
CORRECTED	Low	Execute corrected version
CORRECTED	High	Request principal approval
UNCERTAIN	Any	Request principal approval

3. Agent Registration

3.1 Registration Request

Agents MUST register with QWED before use:

{
  "agent": {
    "name": "CustomerSupportBot",
    "type": "autonomous",
    "description": "Handles customer inquiries",
    "principal_id": "org_abc123",
    "framework": "langchain",
    "model": "claude-3.5-sonnet"
  },
  "permissions": {
    "allowed_engines": ["math", "fact", "sql"],
    "allowed_tools": ["database_read", "send_email"],
    "blocked_tools": ["database_write", "file_delete"]
  },
  "budget": {
    "max_daily_cost_usd": 100.00,
    "max_requests_per_hour": 1000,
    "max_tokens_per_request": 4096
  },
  "trust_level": "supervised"
}

3.2 Registration Response

{
  "agent_id": "agent_xyz789",
  "agent_token": "qwed_agent_...",
  "status": "active",
  "created_at": "2025-12-20T00:30:00Z",
  "permissions": { ... },
  "budget": { ... }
}

3.3 Agent Types

Type	Description	Trust Level
`supervised`	Human approval for high-risk actions	Low
`autonomous`	Self-executing within limits	Medium
`trusted`	Full autonomy (enterprise only)	High

3.4 Agent Identity

Agents receive a DID-based identity:

did:qwed:agent:<agent_id>

4. Verification Requests

4.1 Agent Verification Request

The context object with conversation_id and step_number is required. The step_number must be a positive integer that increases monotonically within a conversation. QWED uses these fields to enforce replay protection, loop detection, and conversation length limits. See conversation controls for details.

{
  "agent_id": "agent_xyz789",
  "agent_token": "qwed_agent_...",
  "action": {
    "type": "execute_sql",
    "query": "SELECT * FROM customers WHERE status = 'active'",
    "target": "production_db"
  },
  "context": {
    "conversation_id": "conv_123",
    "step_number": 5,
    "user_intent": "Get list of active customers"
  },
  "options": {
    "require_attestation": true,
    "risk_threshold": "medium"
  }
}

4.2 Verification Response

{
  "decision": "APPROVED",
  "verification": {
    "status": "VERIFIED",
    "engine": "sql",
    "risk_level": "low",
    "checks_passed": [
      "no_destructive_operations",
      "no_sensitive_columns",
      "schema_valid"
    ]
  },
  "attestation": "eyJhbGciOiJFUzI1NiIs...",
  "budget_remaining": {
    "daily_cost_usd": 89.50,
    "hourly_requests": 42
  }
}

4.3 Decision Types

Decision	Meaning	Agent Action
`APPROVED`	Safe to execute	Proceed
`DENIED`	Verification failed	Abort + log
`CORRECTED`	Fixed version available	Use corrected
`PENDING`	Requires human approval	Wait
`BUDGET_EXCEEDED`	Limits reached	Abort

5. Tool Verification

5.1 Tool Call Request

Before an agent calls an external tool:

{
  "agent_id": "agent_xyz789",
  "tool_call": {
    "tool_name": "send_email",
    "parameters": {
      "to": "user@example.com",
      "subject": "Your order status",
      "body": "Your order #12345 has shipped..."
    }
  },
  "justification": "User requested order status update"
}

5.2 Tool Risk Assessment

{
  "tool_name": "send_email",
  "risk_assessment": {
    "base_risk": "medium",
    "factors": [
      {"factor": "external_communication", "weight": 0.3},
      {"factor": "pii_in_content", "weight": 0.5}
    ],
    "final_risk": "medium",
    "requires_approval": false
  },
  "policy_checks": [
    {"policy": "no_pii_leakage", "passed": true},
    {"policy": "rate_limit", "passed": true}
  ]
}

5.3 Tool Registry

{
  "tools": [
    {
      "name": "database_read",
      "risk_level": "low",
      "requires_verification": true,
      "verification_engine": "sql"
    },
    {
      "name": "database_write",
      "risk_level": "critical",
      "requires_verification": true,
      "requires_approval": true,
      "verification_engine": "sql"
    },
    {
      "name": "execute_code",
      "risk_level": "critical",
      "requires_verification": true,
      "verification_engine": "code",
      "sandbox_required": true
    }
  ]
}

6. Budget & Limits

6.1 Budget Schema

{
  "budget": {
    "cost": {
      "max_daily_usd": 100.00,
      "max_per_request_usd": 1.00,
      "current_daily_usd": 10.50
    },
    "requests": {
      "max_per_hour": 1000,
      "max_per_day": 10000,
      "current_hour": 42,
      "current_day": 350
    },
    "tokens": {
      "max_per_request": 4096,
      "max_daily": 1000000,
      "current_daily": 50000
    },
    "tools": {
      "max_calls_per_hour": 100,
      "high_risk_calls_remaining": 5
    }
  }
}

6.2 Budget Enforcement

┌─────────────────────────────────────────────────────────────┐
│                  BUDGET CHECK FLOW                          │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Request ──▶ [Check Cost] ──▶ [Check Rate] ──▶ [Execute]   │
│                   │                │                        │
│                   ▼                ▼                        │
│              Exceeded?         Exceeded?                    │
│                   │                │                        │
│              ┌────┴────┐     ┌────┴────┐                   │
│              │  DENY   │     │  DENY   │                   │
│              │  +429   │     │  +429   │                   │
│              └─────────┘     └─────────┘                   │
│                                                             │
└─────────────────────────────────────────────────────────────┘

6.3 Budget Response

{
  "decision": "BUDGET_EXCEEDED",
  "error": {
    "code": "QWED-AGENT-BUDGET-001",
    "message": "Daily cost limit exceeded",
    "details": {
      "limit": 100.00,
      "current": 102.50,
      "reset_at": "2025-12-21T00:00:00Z"
    }
  }
}

7. Audit Trail

7.1 Activity Log Schema

Every agent action is logged:

{
  "activity_id": "act_abc123",
  "agent_id": "agent_xyz789",
  "timestamp": "2025-12-20T00:30:00Z",
  "action": {
    "type": "tool_call",
    "tool": "database_read",
    "parameters": { ... }
  },
  "verification": {
    "status": "VERIFIED",
    "engine": "sql",
    "latency_ms": 45
  },
  "decision": "APPROVED",
  "execution": {
    "success": true,
    "result_hash": "sha256:..."
  },
  "cost": {
    "usd": 0.05,
    "tokens": 150
  },
  "attestation_id": "att_xyz789"
}

7.2 Audit Query API

GET /agents/:agent_id/activity?from=2025-12-01&to=2025-12-20

Response:

{
  "agent_id": "agent_xyz789",
  "period": {
    "from": "2025-12-01T00:00:00Z",
    "to": "2025-12-20T00:00:00Z"
  },
  "summary": {
    "total_actions": 15420,
    "approved": 15200,
    "denied": 180,
    "corrected": 40,
    "total_cost_usd": 850.00
  },
  "activities": [ ... ]
}

7.3 Compliance Export

GET /agents/:agent_id/compliance-report?format=pdf

8. Trust Levels

8.1 Trust Level Definitions

Level	Description	Verification	Approval
0: Untrusted	No autonomous actions	All	All
1: Supervised	Low-risk autonomous	High-risk	High-risk
2: Autonomous	Most actions autonomous	Critical only	Critical only
3: Trusted	Full autonomy	None	None

8.2 Trust Elevation

Agents can request trust elevation:

{
  "agent_id": "agent_xyz789",
  "request": "trust_elevation",
  "from_level": 1,
  "to_level": 2,
  "justification": "30 days of safe operation",
  "evidence": {
    "days_active": 30,
    "total_actions": 50000,
    "denied_actions": 50,
    "denial_rate": 0.001,
    "attestations": 50000
  }
}

8.3 Trust Degradation

Automatic trust reduction on violations:

Violation	Penalty
Security policy violation	-2 levels
Repeated denials (>10%)	-1 level
Budget abuse	-1 level
Principal complaint	Suspend

9. Runtime Hardening

New in v1.1.0

9.1 Action Context Requirements

All verification requests MUST include a context with:

Field	Type	Required	Constraints
`conversation_id`	string	Yes	Non-empty identifier for the conversation session
`step_number`	integer	Yes	Must be >= 1 and monotonically increasing within a conversation
`user_intent`	string	No	Human-readable description of intent

The maximum number of steps per conversation is 50. Exceeding this limit triggers QWED-AGENT-LOOP-001.

9.2 Replay Detection

The runtime tracks the highest committed step number per (agent_id, conversation_id) pair. A verification request with a step_number less than or equal to the last committed step is rejected as a replay (QWED-AGENT-LOOP-002). Step numbers are only committed when the action decision is APPROVED or PENDING. Denied actions do not advance the conversation state, allowing the agent to retry the same step number with a different action.

9.3 Repetitive Loop Detection

Actions are fingerprinted using a deterministic JSON serialization of:

action_type
query
code
target
parameters

If the same fingerprint appears more than 2 consecutive times, the action is blocked with QWED-AGENT-LOOP-003. The repeat counter resets when a different action is submitted. Action parameters MUST be deterministic JSON-compatible values (strings, numbers, booleans, nulls, arrays, and objects with string keys). Non-finite floats (NaN, Infinity) and non-string dictionary keys are rejected.

9.4 In-Flight Reservations

To prevent race conditions in concurrent environments, the runtime uses a reservation system:

When a verification request begins processing, the step number is reserved
Concurrent requests for the same step are rejected with QWED-AGENT-LOOP-002
If the action is denied, the reservation is released — the step can be retried
If the action is approved or pending, the reservation is committed — the step is permanently consumed

9.5 Budget Denial Semantics

Budget check failures (QWED-AGENT-BUDGET-001, QWED-AGENT-BUDGET-002) do not consume the conversation step. The in-flight reservation is released so the agent can retry the same step number after the budget resets.

9.6 Fail-Closed Rate Limiting

When the Redis backend is unavailable, the sliding window rate limiter fails closed (denies all requests) rather than failing open. This prevents uncontrolled access during infrastructure failures. When Redis is entirely absent at process startup, a local in-memory fallback limiter is used.

9.7 Environment Integrity

On server startup, the runtime MUST verify environment integrity (via StartupHookGuard) before initializing the database. A compromised environment causes the server to abort startup with a RuntimeError.

9.8 Timing-Safe Authentication

Agent token verification MUST use constant-time comparison (hmac.compare_digest) to prevent timing side-channel attacks.

10. Implementation Guidelines

9.1 SDK Integration

from qwed_sdk import QWEDAgentClient

# Register agent
agent = QWEDAgentClient.register(
    name="MyAgent",
    principal_id="org_123",
    permissions={
        "allowed_engines": ["math", "sql"],
        "allowed_tools": ["database_read"]
    },
    budget={
        "max_daily_cost_usd": 50.00
    }
)

# Before executing any action
result = agent.verify_action({
    "type": "execute_sql",
    "query": "SELECT * FROM users"
})

if result.decision == "APPROVED":
    # Safe to execute
    execute_query(result.verified_query)
    
    # Log with attestation
    agent.log_execution(
        action_id=result.action_id,
        success=True,
        attestation=result.attestation
    )

9.2 LangChain Integration

from langchain.agents import AgentExecutor
from qwed_sdk.langchain import QWEDVerificationCallback

# Wrap agent with QWED verification
agent_executor = AgentExecutor(
    agent=my_agent,
    tools=my_tools,
    callbacks=[QWEDVerificationCallback(
        agent_id="agent_xyz789",
        agent_token="qwed_agent_..."
    )]
)

# All tool calls automatically verified
result = agent_executor.run("Get customer data")

9.3 CrewAI Integration

from crewai import Agent, Task, Crew
from qwed_sdk.crewai import QWEDVerifiedAgent

# Wrap agents with QWED
verified_agent = QWEDVerifiedAgent(
    agent=researcher,
    qwed_settings={
        "verify_all_tools": True,
        "require_attestation": True
    }
)

Appendix A: Error Codes

Code	Description
`QWED-AGENT-001`	Agent not registered
`QWED-AGENT-002`	Invalid agent token
`QWED-AGENT-003`	Agent suspended
`QWED-AGENT-004`	Tool not allowed
`QWED-AGENT-005`	Verification failed
`QWED-AGENT-CTX-001`	Missing required action context (`conversation_id` and `step_number`)
`QWED-AGENT-CTX-002`	Invalid `step_number` (must be >= 1)
`QWED-AGENT-LOOP-001`	Conversation step limit exceeded (max 50 steps)
`QWED-AGENT-LOOP-002`	Replay or out-of-order action step detected
`QWED-AGENT-LOOP-003`	Repetitive action loop detected (max 2 consecutive identical actions)
`QWED-AGENT-BUDGET-001`	Daily cost exceeded
`QWED-AGENT-BUDGET-002`	Hourly rate exceeded
`QWED-AGENT-BUDGET-003`	Token limit exceeded
`QWED-AGENT-TRUST-001`	Insufficient trust level
`QWED-AGENT-TRUST-002`	Action requires approval

Appendix B: HTTP Endpoints

Endpoint	Method	Description
`/agents/register`	POST	Register new agent
`/agents/:id`	GET	Get agent details
`/agents/:id/verify`	POST	Verify agent action
`/agents/:id/tools/:tool`	POST	Verify tool call
`/agents/:id/activity`	GET	Get activity log
`/agents/:id/budget`	GET	Get budget status
`/agents/:id/trust`	POST	Request trust change

Architecture

Concepts

Specifications

Integration Patterns

Contributing

​Table of Contents

​1. Introduction

​1.1 Purpose

​1.2 Problem Statement

​1.3 Solution

​1.4 Terminology

​2. Agent Verification Model

​2.1 Verification Flow

​2.2 Verification Types for Agents

​2.3 Decision Matrix

​3. Agent Registration

​3.1 Registration Request

​3.2 Registration Response

​3.3 Agent Types

​3.4 Agent Identity

​4. Verification Requests

​4.1 Agent Verification Request

​4.2 Verification Response

​4.3 Decision Types

​5. Tool Verification

​5.1 Tool Call Request

​5.2 Tool Risk Assessment

​5.3 Tool Registry

​6. Budget & Limits

​6.1 Budget Schema

​6.2 Budget Enforcement

​6.3 Budget Response

​7. Audit Trail

​7.1 Activity Log Schema

​7.2 Audit Query API

​7.3 Compliance Export

​8. Trust Levels

​8.1 Trust Level Definitions

​8.2 Trust Elevation

​8.3 Trust Degradation

​9. Runtime Hardening

​9.1 Action Context Requirements

​9.2 Replay Detection

​9.3 Repetitive Loop Detection

​9.4 In-Flight Reservations

​9.5 Budget Denial Semantics

​9.6 Fail-Closed Rate Limiting

​9.7 Environment Integrity

​9.8 Timing-Safe Authentication

​10. Implementation Guidelines

​9.1 SDK Integration

​9.2 LangChain Integration

​9.3 CrewAI Integration

​Appendix A: Error Codes

​Appendix B: HTTP Endpoints

Table of Contents