Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Agentic AI

AWS moves agent guardrails outside the model with AgentCore Policy

On December 2, at its re:Invent conference, Amazon Web Services announced two preview features for Amazon Bedrock AgentCore, its service for running AI agents in production: Policy, which sets boundaries on what an agent can do, and Evaluations, which monitors the quality of an agent’s behaviour on live traffic. AWS also introduced what it calls frontier agents, autonomous agents for software development, security testing and incident response.

The design choice behind Policy is the interesting part. As VentureBeat put it, the rules sit outside the agent, between the agent and the tools it calls, rather than inside the prompt.

/ 01What AWS announced

According to the AWS News Blog, Policy applies rules outside the agent’s reasoning loop and intercepts tool calls passing through AgentCore Gateway before they run. Policies can be written in natural language or in Cedar, an open-source policy language. A policy engine can either enforce decisions or only log them, so teams can test a rule before switching it on. VentureBeat’s example was a refund agent limited to $100, with larger requests escalated to a human; when the agent breaks a rule, it is sent back to reconsider.

David Richardson, AWS vice president for AgentCore, framed this partly as a defence against manipulated inputs. “There are always these prompt injection attacks,” he told VentureBeat; a check that sits outside the model is not talked out of its rules by a poisoned document.

Evaluations offers built-in evaluators for dimensions including correctness, faithfulness, helpfulness, tool selection accuracy, safety, goal success rate and context relevance, alongside custom evaluators. AWS also added an episodic memory strategy, intended to help agents learn from past experiences, and bidirectional streaming for voice agents.

/ 02Why it matters

For two years, the default way to make an agent behave was to tell it to behave: system prompts, instructions, examples. That works until it does not. An instruction in a prompt is one more piece of text the model weighs against everything else in its context, including text an attacker put there.

Moving the rule out of the prompt and into a deterministic checkpoint changes the guarantee. The agent can still propose anything; the gateway decides what executes. That is how access control has worked in software for decades, and it is a sign of maturity that agent platforms are adopting it.

/ 03The knowledge-graph angle

A policy engine can only be as good as the facts it checks against. “Refunds over $100 need approval” is simple. Real enterprise rules are rarely that self-contained: this supplier is on a restricted list because of its parent company; this patient cannot receive that drug because of a documented interaction; this dataset may not leave a given jurisdiction. Checking those rules means knowing how entities relate, which is what a knowledge graph records.

The pairing is natural. Policies express what is allowed; a governed graph supplies the facts that decide whether a specific action meets that policy; and a deterministic check produces the same verdict every time, with a record of which facts and which rule were used. That record is what an auditor, or an operator investigating an incident, needs to see.

Evaluations complement this, but they answer a different question. Evaluators score how well an agent behaved, often using another model as the judge. That is useful for spotting drift and regressions across many conversations. It is not the same as being able to say, for one specific decision, exactly why it was made. Enterprises will want both: statistical quality monitoring across the fleet, and a deterministic, inspectable record for the individual actions that carry real consequences.

Telling an agent the rules is not the same as enforcing them. Enforcement belongs outside the model, grounded in facts you can inspect.

/ 04What to do

  • Inventory agent actions, not just agent prompts. Every tool call with side effects should have an explicit rule.
  • Start in log-only mode, as AWS suggests, to see how often real traffic would trip a policy before enforcing it.
  • Make the facts behind policies explicit. If a rule depends on relationships between entities, those relationships need to live in a governed store, not in someone’s head or a prompt.
Aktiver’s take — Enterprise data

Putting the rule between the agent and its tools is the right call, and it should become the default shape for any agent with side effects. The gap is the rule’s inputs: a Cedar policy can say “block payments to restricted suppliers,” but something has to know which suppliers are restricted, and why.

In a knowledge graph, that “why” is a path: supplier, owned by a parent company, parent named in a sanctions notice, notice dated and sourced. A policy check that reads the path instead of a flat allow-list can return its verdict together with the facts it used. With Knowledge Builder, agents enforce semantic reasoning over that graph, so a tool call that does not fit the model fails loudly rather than slipping through. Reasoning is deterministic, so the same proposed action gets the same verdict on replay, and an incident reviewer can ask a precise question: which rule, and which edge in the graph, allowed this call on this date.

How we’d model it

Agent tool callPolicyRestricted supplierSupplierParent companyPolicy decisionGraph factevaluatedAgainstreferencesownedBycitesEvidenceAgent toolcallPolicyRestrictedsupplierSupplierParentcompanyPolicydecisionGraph factevaluatedAgainstreferencesownedBycitesEvidence
  • Agent tool call —evaluatedAgainst→ Policy
  • Policy —references→ Restricted supplier
  • Supplier —ownedBy→ Parent company
  • Policy decision —citesEvidence→ Graph fact

See where this applies →

Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading