Glossary
›
Compiled vs. configured agents

Compiled vs. configured agents

Glossary
Also known as:
Derived vs. authored agents, data-built vs. prompt-built agents

What is the difference between compiled and configured agents?

Compiled and configured agents are two ways of building an AI agent: a configured agent's behavior is written by hand through prompts, intents, rules and knowledge-base articles, while a compiled agent's behavior is derived from an organization's recorded conversations, which are analyzed and turned into a structured flow that the agent runs.

The two words borrow from software. Configuring means setting a system up by hand, option by option. Compiling means translating source material into something a machine can execute, and here the source material is the record of how conversations actually went. Most agents in production use some of both, so the distinction is best read as a question: where does the agent's behavior come from?

How compiled and configured agents are built

The approaches differ most in their starting material and in what happens as the business changes.

AspectConfigured agentCompiled agent
Source of behaviorPrompts, scripts, intents and documents written by a teamRecorded calls, chats and the outcomes that followed them
What it capturesThe process as documentedThe process as performed, including what the most successful people do differently
Effort to startLow at first; grows with every intent, edge case and exceptionHigher up front: recordings must be gathered, analyzed and reviewed
UpkeepManual edits as products and policies changeRe-derived as new conversations accumulate, then reviewed again
Best fitNew products, simple tasks, little recorded historyEstablished, high-volume conversations where results vary by who handles them

In a configured build, a team writes the prompts, defines what the agent should recognize, connects knowledge sources and adds rules as gaps appear. In a compiled build, recordings and transcripts are analyzed to find the structure of successful conversations: the sequence of questions, the decision points, the objections and the responses that preceded good outcomes. People review and adjust that structure, approve it, and the agent runs it. Neither approach removes the need for review; they differ in what reviewers start from.

Why the distinction matters

Hand-written instructions are more fragile than they look. In a study of prompt formatting by Sclar and colleagues, small changes to how a prompt was laid out, with its meaning unchanged, produced performance differences of up to 76 accuracy points on one open model. An agent whose behavior lives mainly in prose instructions can shift in ways that are hard to predict and hard to review.

The second reason is evidence. Process documents describe how work is supposed to go. Recordings show how it went, including the moves that separate the people with the best results from everyone else, which are often never written down. A configured agent reproduces the documentation; a compiled one can reproduce the practice, once that practice has been reviewed.

Where each approach is used

  • Configured, for new territory. New products, new markets and processes with little recorded history.
  • Configured, for simple service. Question answering and routing, where a knowledge base and a few rules are enough.
  • Compiled, for established volume. Applications, quotes, renewals and collections, where thousands of recorded conversations exist and outcomes vary by who handled them.
  • Compiled, for regulated sequences. Conversations where the order of steps and the exact statements matter, and a reviewable structure is easier to approve than a long prompt.
  • Both, in most production agents. The main flow is compiled from recordings, while new offers, policy wording and seasonal changes are configured on top.

A team without recordings, or with very low volume, should configure first and collect data. Compiling needs enough examples of each situation to separate what works from what happened once.

What a good build looks like, either way

  • Traceable behavior. Every rule or step can be traced to a source: a document, a decision or a set of recorded conversations.
  • Approval before launch. Compliance and the business owner sign off on the structure and the fixed statements.
  • Replay against history. The agent is run against past conversations to see how it would have handled them.
  • Locked wording for regulated statements. Disclosures are fixed text, not generated.
  • A defined update path. It is clear who changes what, how changes are tested and how they are undone.

How to measure compiled and configured agents

  • Outcome rate against a baseline. The result each approach achieves on comparable conversations.
  • Coverage. The share of real conversations the agent handles without falling back to a generic response or a person.
  • Time to first launch. From project start to the first live conversation.
  • Maintenance load. Hours spent each month on edits, fixes and new cases.
  • Consistency. Whether customers in the same situation receive the same treatment.

Regulatory questions each build method raises

How an agent is built does not change the rules it must follow, but each method raises its own questions. These are the main US considerations; other jurisdictions and regulated sectors have their own.

  • Using recorded conversations. California's privacy regulations require that personal information be used only in ways reasonably necessary and proportionate to the purpose for which it was collected, or a compatible one. Building an agent from customer recordings should fit within what customers were told.
  • Repeating past conduct. Practices found in old conversations, such as inconsistent offers to similar customers, become systematic once compiled. Fair lending law forbids treating credit applicants differently on a prohibited basis, so compiled behavior should be checked for that before release.
  • Required statements. Whichever method is used, statements the law requires, such as a debt collector's disclosure that it is attempting to collect a debt, must be delivered as required and cannot depend on how a prompt is phrased.

Frequently asked questions

Which is better, compiled or configured?

Neither in general. Configured is faster to start and suits new or simple processes. Compiled suits established conversations with a large recorded history, where the goal is to reproduce what already works.

Can a compiled agent handle a new product?

The parts of the conversation that carry over, such as qualifying questions and objection handling, often transfer. Product-specific content is configured until enough new conversations exist to learn from.

Does compiling copy individual employees?

It captures patterns that recur across many successful conversations, such as sequences and responses, not a person's voice or identity. Recordings are typically anonymized before analysis.

Is prompt engineering a form of configuration?

Yes. Writing and refining prompts is configuration by hand, however long the prompts and however capable the model.

Can a configured agent become compiled later?

Yes. Once it has handled enough conversations with recorded outcomes, those records can be analyzed and used to restructure it.

Compiled from the conversations you already have.

Encore AI's Interaction Mining analyzes your recorded calls and chats to find how your top performers win, compiles that into an executable flow graph governed as a playbook, and runs it as AI agents across voice, chat, SMS, WhatsApp, email, IVR, IVA and agentic landing pages.

Scale your best to be the best.

Book a Demo