Skip to main content

Polidex vs. System Prompts

System prompts influence agent behavior. Polidex enforces it.

That gap — between influence and enforce — is the whole argument.


Why System Prompts Fade as Policy

A system prompt is a text instruction. The agent reads it at the start of a conversation and does its best to follow it. On routine cases in short conversations, that works well enough. On complex cases — a frustrated customer, a long back-and-forth, an ambiguous exception — the instruction competes with everything else in the context window for the model's attention.

Transformer models are designed to weight recent context more heavily than static instructions. As a conversation grows longer, the system prompt fades. What was an authoritative rule at turn 1 is, by turn 25, one signal among many. Your refund policy instruction doesn't disappear — it just carries less weight exactly when the stakes are highest.

This is not a bug in any particular model. It is how the architecture works. You cannot engineer around it with better prompt writing.

The only fix is moving policy enforcement out of the context window entirely.


The Structural Difference

System Prompt as PolicyPolidex
EnforcementProbabilistic — agent interprets the instructionStructural — agent cannot act without an authorized decision
Version historyNoneEvery rule versioned with effective dates
Policy change governanceNoneVersioned, attributed, and auditable
Audit trailNoneEvery decision tied to a policy version, timestamp, and authorization path
Policy ownerWhoever has edit access to a text fileBusiness operator through a governed interface
Policy updateEdit the file, manually push to every agent configurationUpdate once — propagates immediately to every agent that calls the layer

What Structural Enforcement Means

Think about how payment authorization works.

A merchant cannot charge a card by deciding to skip the payment network. The card network is a required technical intermediary. The charge either clears through the network or it doesn't go through. No amount of merchant intent changes that.

Decision tokens work the same way. When your agent needs to issue a refund, apply a credit, or grant an exception, it calls Polidex first. Polidex evaluates the relevant policy — what version is active, what rules apply, whether approval is required — and returns a resolved decision with a signed token. The downstream action requires that token. No token, no action.

The agent isn't reading a rule and deciding whether to follow it. The enforcement is in the architecture, not in the agent's behavior.


The Update Problem

Your refund policy changes. With system prompts, someone edits a text file. There is no approval record. No version history. No guarantee every agent instance got the update simultaneously. No way to answer "what rule was active on March 14th at 2:17pm?" without forensic work — and even then, probably not.

With policy versioning, the old rule is archived, not overwritten. Every decision references the policy version that authorized it. The update is governed — who changed it, when, with what authorization. You can roll back if a change had unintended consequences.

That record is what an audit asks for. "It was in the system prompt, and here's the current file" does not answer the question.


The Decision

If your agents are making policy decisions — refunds, exceptions, escalations, entitlement calls — there are two architectures available.

One puts policy in text that the agent reads and interprets. The other puts policy in a layer the agent is required to call. The first is easier to set up. The second is the only one that produces consistent, auditable decisions at agent speed.

For the full technical analysis of why system prompts fail as policy — attention decay, the three structural failure modes, and the OWASP guidance on enforcement — read System Prompts Aren't Policy.

Ready to talk?

Tell us how we can help.

Get in Touch