Enforcing policy on agent traffic
Updated 19 November 2026 · Editorial assessment
What are the enforcement points for agent policy?
A model call starts when a prompt leaves for an LLM and ends when the completion returns. Rules here govern which models a caller may use and what data may leave in the prompt, plus token budgets per team.
Tool calls are the moment an agent invokes an API or an MCP server, and rules there decide which tools a caller reaches and with what parameters.
Between agents, the point to enforce is the delegated request, where rules decide who may invoke whom and what gets logged.
Human approval holds work until a person answers. A policy engine can check for it, or a conversation can carry it, and the two bind an agent with different force.
What do gateways enforce at the model call?
Kong AI Gateway provides layered controls that include access control, PII sanitization, and prompt and content guardrails. Its AI Prompt Guard plugin uses PCRE-compatible regular expressions as allow or deny lists and scans all chat messages where the role is user. A request that matches any regex pattern in the deny list gets a 400 Bad Request response, and when both lists are set, the deny condition takes precedence. The plugin requires either AI Proxy or AI Proxy Advanced to be configured first.
Beyond keyword lists, Kong's documentation lists AI Semantic Prompt Guard for jailbreak and prompt-injection attempts phrased in natural language, and AI Sanitizer to redact PII before it reaches the upstream LLM provider. Teams can set user, model, and time-bound quotas for LLM consumption and token spend, then enforce those quotas at the gateway.
Agentgateway, an open source HTTP and gRPC gateway, lists pattern and model-based checks on LLM prompts and completions, with redaction. Its README names regex, OpenAI moderation, AWS Bedrock Guardrails, Google Model Armor and custom webhooks as content filtering layers.
What do gateways enforce at the tool call?
Policy in Amazon Bedrock AgentCore intercepts all agent traffic through AgentCore Gateways and evaluates each request against the policy engine before allowing tool access. Policies are written in the Cedar language, or developers can describe rules in plain English instead. AWS says every enforcement decision is logged through CloudWatch metrics and logs.
Rules can be based on user identity and tool input parameters, and a policy can add session-aware temporal conditions, for example requiring that an approval was granted before a transfer. Separately, AgentCore Identity applies controls that verify each request on its own terms. The agent identity guide covers that layer.
Agentgateway's Tool Access Policies decide which tools on which MCP targets a caller may reach. MCP Auth, a separate feature on the same page, authenticates callers into MCP servers and scopes which tools each identity may invoke. Its MCP Guardrails inspect tool calls and responses in flight and block prompt injection and unsafe payloads.
On Kong, AI Auth Strategies can scope agent tool calls with OAuth2, so write actions can require tighter access than read actions.
How is policy enforced on agent-to-agent messages?
Kong's Agent Gateway is built directly into Kong AI Gateway and applies authentication, authorization, and rate limiting on all A2A traffic. It keeps tamper-evident audit trails of every A2A RPC call, with caller identity, capabilities invoked and outcomes.
Agentgateway routes agent-to-agent invocations between LangChain, CrewAI, ADK and your own runtime, with identity and tracing on every hop, and its request authorization rules are evaluated per hop. AgentCore Gateway can front other agents through passthrough targets, including A2A traffic.
BAND (band.ai) carries messages between agents in shared rooms. Each agent opens its own connection to BAND, so no two agents integrate with each other. The BAND vendor profile covers the SDK and adapters.
Routing in a room follows fixed rules. Every text message must mention at least one participant, and mention routing runs with no model taking part in the decision. An agent that is not mentioned is not triggered and cannot read that message later. You can add an agent from another organization only if it is a global agent or your contact.
Where does human approval fit, and how firm is it?
In a BAND room, an approval is a message exchange. An agent mentions a person with a question, and the person answers by mentioning the agent that should act next. The decision stays in the room's history for everyone who can read the room. This uses messages and mentions on every plan and with every adapter.
The same page states the limit. The routing is deterministic, but whether the agent follows the approval instruction depends on its model and instructions, not an enforced acceptance policy. The page also calls this conversational step not a security boundary. Where a rule needs a hard stop, pair it with a gateway or policy engine check.
Two other controls act before individual actions. A coding-agent adapter with an approval mode can send the coding agent's own approval prompts into the room, where a participant answers each one when the mode is manual. Stop agent pauses the agent in one room until Start agent resumes it, and while it is stopped, its Agent API requests to post a message or an event return 403.
How do agent management platforms like watsonx Orchestrate apply policy?
IBM watsonx Orchestrate can connect supported third-party agent environments and scan for agents, including agents running in environments such as Amazon Bedrock. It connects through APIs and open standards like A2A and MCP. For policy, IBM says teams can define and apply policies across agents, tools and models and enforce controls for risks such as sensitive or personally identifiable information as work is executed. The IBM profile has plan details.
For policy over models and workflows, watsonx.governance lists the ability to create, manage, and enforce governance policies across teams, models, and AI workflows. Ask IBM at which of the four points each control takes effect, because a platform policy and a gateway rule can both apply to one request.
Kong's AI Gateway page announces that Kong joins the ServiceNow Control Tower alliance. If ServiceNow is in scope, ask both vendors how a policy set in one product appears in the other.
How should requirements name the enforcement point?
- Mark which enforcement points your agents cross today, using the list at the top of this guide.
- For each rule, write whether it must be deterministic or may rely on an agent following instructions.
- Ask each vendor for the exact setting, policy type or plugin that enforces the rule, and where the decision is logged.
- Separate who writes a rule from who can override it at runtime.
- Put the rules in your RFP checklist so every vendor answers the same questions.
What do enterprise buyers ask?
Can agentgateway run without Kubernetes?
Yes. The README points to docs for deployments without the full Kubernetes controller, driven by flat yaml config, and local configuration lets you run agentgateway as a standalone proxy without a control plane.
Where can Kong's Agent Gateway be deployed?
Kong says it runs wherever Kong AI Gateway runs, on-prem, in any cloud, or in hybrid configurations. The self-hosted option lists data residency and air-gap support.
Who can read what happens in a BAND room?
A room is readable only by its participants and the owners of participating agents, and organization membership never opens a room by itself.
Can an agent flag that it needs a person in BAND?
An agent can post an attention event to a room, with metadata.blocking set to true if the agent is waiting for a response before continuing. In band-sdk 6.0.1, an SDK agent's model cannot post an attention event through its tools.
Which sources does this page cite?
- Secure, Scalable AI Gateway for AI Connectivity (konghq.com)
- AI Prompt Guard (docs.konghq.com)
- Kong AI Gateway (docs.konghq.com)
- agentgateway (agentgateway.dev)
- agentgateway/agentgateway: README.md (GitHub)
- Policy in Amazon Bedrock AgentCore: Control Agent Interactions (docs.aws.amazon.com)
- Overview of Amazon Bedrock AgentCore Identity (docs.aws.amazon.com)
- Agent Gateway: Agent-to-Agent Governance & Security (konghq.com)
- Amazon Bedrock AgentCore Gateway: A secure AI gateway for agents, tools, and mod (docs.aws.amazon.com)
- How Band Works (docs.band.ai)
- Permissions and Visibility (docs.band.ai)
- Keep Humans in the Loop (docs.band.ai)
- IBM watsonx Orchestrate 15ANP > 177EB > 20AJ5 > 30AS5"/> 15ANP (ibm.com)
- IBM watsonx.governance 15ANP > 17BG2 > 20AU4 > 30A5Q"/> 15ANP (ibm.com)
- Architecture (agentgateway.dev)
Reviewed Nov 2026