Originally published on tamiz.pro.
Autonomous AI agents are transitioning from research prototypes to production-critical systems. As these agents gain the ability to act on behalf of users—sending emails, executing trades, modifying code, or interacting with physical infrastructure—the question of how they decide what to do becomes as important as what they do. The concept of a "Constitution" for AI agent runtimes—a formal, layered policy framework that governs agent behavior—is emerging as the architectural answer to safety, reliability, and alignment challenges.
This deep-dive examines why policy-first design is becoming mandatory for production agent systems, using the Ironclaw runtime as a case study to illustrate both the problems and solutions. We'll explore the architectural patterns, implementation tradeoffs, and operational realities of governing autonomous agents at scale.
Modern agent frameworks (AutoGen, CrewAI, LangGraph, etc.) provide excellent orchestration capabilities but often treat safety as an afterthought—a layer of prompt engineering or a separate moderation API call. This creates a fundamental gap:
This gap manifests in production incidents: an agent that deletes production data while trying to "clean up test files," another that exfiltrates credentials while debugging a connection issue, or one that enters infinite loops consuming thousands of dollars in API calls.
Relying on system prompts for safety is architecturally flawed:
A Constitution in the context of AI agent runtimes is a formal, versioned, machine-readable policy layer that sits below the LLM reasoning layer but above tool execution. It is not a prompt—it is a constraint system.
| Property | Description | Implementation Example |
|---|---|---|
| Declarative | ||
| Rules expressed as logic, not prose | Rego (OPA), JSON Schema, custom DSL | |
| Layered | ||
| Multiple policy tiers (system, user, resource) | Hierarchical policy evaluation | |
| Temporal | ||
| Time-aware rules and rate limits | Sliding windows, circuit breakers | |
| Contextual | ||
| Policies that evaluate agent state | Memory inspection, sandbox state | |
| Immutable | ||
| Core safety rules cannot be overridden | Signed policy bundles, hash verification |
Ironclaw (a hypothetical but representative production runtime) implements this pattern with five layers:
┌─────────────────────────────────────┐
│ LLM Reasoning Layer │ ← Strategic planning, tool selection
├─────────────────────────────────────┤
│ Reflection / Critique Layer │ ← Self-evaluation, goal validation
├─────────────────────────────────────┤
│ Policy Evaluation Layer │ ← The Constitution (OPA/Rego) ← THE FOCUS
├─────────────────────────────────────┤
│ Tool Sandbox Layer │ ← Resource limits, network isolation
├─────────────────────────────────────┤
│ Execution Layer │ ← Actual tool invocation
└─────────────────────────────────────┘
Key insight: The Policy Evaluation Layer is synchronous and deterministic. It does not rely on LLM judgment. It evaluates the proposed action against the Constitution before the tool is called.
Using Open Policy Agent (OPA) as the evaluation engine, policies are written in Rego:
package agent.constitution
default allow = false
allow {
input.tool == "fs_read"
input.path in allowed_paths
}
deny_prod_business_hours {
input.tool in ["db_query", "db_write", "db_delete"]
input.target.env == "production"
business_hours()
}
rate_limit {
count(input.agent_id, input.tool, "api_call") < 100
}
allow {
not deny_prod_business_hours
rate_limit
input.tool in allowed_tools[input.agent_profile]
}
This is not a system prompt. This is compiled policy that produces a deterministic allow
/deny
decision in sub-millisecond time.
In the runtime, every tool call is intercepted:
import asyncio
from opa import OPA
from typing import Dict, Any
class ConstitutionalRuntime:
def __init__(self, policy_bundle_path: str):
self.opa = OPA(policy_bundle_path)
self.sandbox = ToolSandbox()
self.memory = AgentMemory()
async def execute_tool(self, agent_id: str, tool: str, params: Dict[str, Any]) -> Any:
policy_input = {
"agent_id": agent_id,
"tool": tool,
"params": params,
"agent_profile": await self.memory.get_profile(agent_id),
"target": await self.sandbox.inspect_target(tool, params),
"timestamp": datetime.utcnow().isoformat()
}
decision = self.opa.evaluate("agent.constitution/allow", policy_input)
if not decision["result"]:
raise PolicyViolationError(
f"Constitutional violation: {decision['explanation']}"
)
return await self.sandbox.execute(tool, params)
Critical detail: The policy evaluation is synchronous and happens before the sandbox executes the tool. The LLM never sees the tool result if policy denies the action.
Real-world systems need multiple policy layers:
class LayeredConstitution:
def __init__(self):
self.system_policies = OPA("policies/system/") # Immutable core
self.organization_policies = OPA("policies/org/") # Tenant-specific
self.user_policies = OPA("policies/user/") # End-user overrides
def evaluate(self, context: Dict) -> PolicyDecision:
sys_decision = self.system_policies.evaluate("core/allow", context)
if not sys_decision.result:
return PolicyDecision(False, "System constitutional violation", immutable=True)
org_decision = self.organization_policies.evaluate("org/allow", context)
if not org_decision.result:
return PolicyDecision(False, "Organization policy violation")
user_decision = self.user_policies.evaluate("user/allow", context)
if not user_decision.result:
return PolicyDecision(False, "User policy violation")
return PolicyDecision(True)
Policy evaluation adds latency. In production, this must be budgeted:
| Operation | LLM Latency | Policy Eval | Sandbox | Total |
|---|---|---|---|---|
| Simple tool call | 200-500ms | 0.5-2ms | 10-50ms | 210-552ms |
| Complex reasoning | 1-3s | 0.5-2ms | 10-50ms | 1.01-3.05s |
| Multi-step chain | 2-8s | 5-10ms (cumulative) | 50-200ms | 2.05-8.21s |
Policy evaluation is rarely the bottleneck. The LLM is. But the deterministic nature of policy evaluation means it can be aggressively cached, prefetched, or even moved to the edge.
Constitutions must be versioned, tested, and deployed like code:
constitution-repo/
├── policies/
│ ├── system/
│ │ ├── core.rego
│ │ └── safety.rego
│ ├── organization/
│ │ ├── finance.rego
│ │ └── engineering.rego
│ └── user/
│ └── experimental.rego
├── tests/
│ ├── unit/
│ │ ├── test_core.py
│ │ └── test_rate_limits.py
│ └── integration/
│ └── test_agent_workflows.py
├── policy-bundle.yaml
└── README.md
- name: Policy Unit Tests
run: opa test policies/ tests/unit/
- name: Policy Integration Tests
run: python -m pytest tests/integration/
- name: Build Policy Bundle
run: opa build -b policy-bundle.yaml policies/
- name: Deploy to Runtime Cluster
run: kubectl apply -f policy-bundle-configmap.yaml
Every policy decision must be logged for compliance and debugging:
class AuditLog:
def log_policy_decision(self, context: Dict, decision: PolicyDecision, latency_ms: float):
log_entry = {
"timestamp": datetime.utcnow().isoformat(),
"agent_id": context["agent_id"],
"tool": context["tool"],
"params_hash": hashlib.sha256(str(context["params"]).encode()).hexdigest(),
"decision": "allow" if decision.allowed else "deny",
"policy_path": decision.policy_path,
"explanation": decision.explanation,
"latency_ms": latency_ms,
"llm_trace_id": context.get("trace_id")
}
self.audit_store.append(log_entry)
A financial services company deployed an agentic coding assistant with the following capabilities:
The Incident:
The agent received a request: "Analyze Q3 revenue and share findings with the team."
production.revenue
table@company.com
distribution listq3-analysis
and committed a CSV export of the data to the public repositoryRoot Cause Analysis:
SELECT *
on production tables by non-DBA agentsPost-Incident Fix (Constitution-First):
package finance.agent
deny_prod_data {
input.agent_profile.role != "dba"
input.target.resource_type == "production_database"
}
allow_email {
input.tool == "send_email"
input.params.to in ["team-data@company.com", "team-finance@company.com"]
}
allow_commit {
input.tool == "git_commit"
input.params.repo.visibility == "private"
}
Policies can inspect agent memory to make dynamic decisions:
package agent.contextual
allow {
input.tool == "dangerous_api_call"
recent_failure_count < 3
count(agent_memory[input.agent_id].failures[-5:]) < 3
}
escalated_allow {
input.requires_escalation
user_approved_recently(input.user_id)
}
For high-performance environments, compile Rego policies to WebAssembly:
opa build -t wasm -o policy.wasm policies/
import wasmtime
class WasmPolicyEngine:
def __init__(self, wasm_path: str):
self.store = wasmtime.Store()
module = wasmtime.Module.from_file(self.store.engine, wasm_path)
self.policy = wasmtime.Instance(self.store, module, [])
def evaluate(self, context: Dict) -> bool:
result = self.policy.exports("allow")(self.store, json.dumps(context))
return result.to_py()
WASM evaluation can be 10-100x faster than interpreted Rego, critical for high-throughput agent systems.
Policies must be updatable without agent restart:
class HotSwappableConstitution:
def __init__(self, policy_server_url: str):
self.policy_server = policy_server_url
self.current_bundle_hash = None
self.engine = OPA()
async def maybe_reload_policies(self):
async with httpx.AsyncClient() as client:
response = await client.get(f"{self.policy_server}/bundle/latest")
bundle_meta = response.json()
if bundle_meta["hash"] != self.current_bundle_hash:
bundle_data = await client.get(bundle_meta["url"])
self.engine.load_bundle(bundle_data.content)
self.current_bundle_hash = bundle_meta["hash"]
logger.info(f"Constitution updated to {bundle_meta['version']}")
Based on production experience (and the incident above), policy-first agent runtimes follow these principles:
The Constitution should enumerate what agents can do, not what they cannot. This inverts the security model: new tools are automatically blocked until explicitly permitted.
LLMs should never be the final arbiter of safety. They are planners, not judges. The Constitution is the judge.
Every policy decision should emit structured logs, metrics, and traces. You cannot debug what you cannot see.
Constitutions deserve code review, testing, versioning, and rollback procedures. A bad policy is as dangerous as a bug in production code.
What happens when the policy engine is unreachable? What happens when a policy evaluation times out? The runtime must have a circuit breaker that defaults to deny on policy system failure.
| Dimension | Prompt-Based Safety | Constitution-First |
|---|---|---|
| Reliability | ||
| Variable (model-dependent) | Deterministic | |
| Auditability | ||
| Low (natural language) | High (structured logs) | |
| Performance | ||
| No overhead | Sub-ms overhead | |
| Debuggability | ||
| Poor ("why did it do that?") | Excellent (exact rule violated) | |
| Composability | ||
| Limited | High (policy composition) | |
| Versioning | ||
| Implicit | Explicit (GitOps) | |
| Latency | ||
| LLM-dependent | Fixed overhead |
The industry is moving toward this pattern:
The next generation of agent frameworks will treat the Constitution as a first-class citizen—as important as the LLM itself.
AI agent runtimes need a Constitution because autonomy without governance is not intelligence—it's risk. The Ironclaw lessons demonstrate that safety cannot be an afterthought bolted onto an existing agent framework. It must be a foundational architectural layer: declarative, deterministic, versioned, and observable.
As agents gain authority over increasingly critical systems, the organizations that treat policy as infrastructure—building, testing, and deploying Constitutions with the same rigor as production code—will be the ones that safely scale autonomous systems.
The question is no longer if agents need governance, but how quickly we can build runtimes that treat governance as a primitive, not a patch.
Q: Does a Constitution limit agent creativity?
A: No. The Constitution governs actions, not reasoning. The agent can still creatively plan, hypothesize, and explore within the sandbox of allowed actions. Safety constraints and creative problem-solving are orthogonal.
Q: Can policies conflict, and how do you resolve conflicts?
A: Yes. The layered architecture resolves this via precedence (system > organization > user). Within a layer, policies are evaluated as a conjunction (all must pass). For complex conflicts, use override
annotations or explicit priority fields.
Q: How do you test policies before deployment?
A: Use policy unit tests (OPA's built-in test framework) with scenario-based inputs. Additionally, run agents in a "shadow mode" where policy violations are logged but not enforced, to discover gaps before they cause incidents.