> ## Content Index
> Fetch the complete content index at: https://cms.syntrigen.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# What “production” means for an agent
- URL: https://cms.syntrigen.com/what-production-means-for-an-agent/
- Published: 2026-04-15T09:00:00.000Z
- Updated: 2026-08-20T12:48:35.000Z
- Description: Eleven questions that separate a system from a demo. None of them are about model quality, and most projects that stall have not answered more than three.
- Author: Syntrigen
- Tags: AI in production, Platform

"It works" and "it is in production" are separated by a body of unglamorous work that rarely appears in a plan. Below is the list we actually walk through. It is not a maturity model and there are no levels — each item is either answered or it is not.

## 1\. What happens on the second attempt?

A tool call times out. The agent retries. Does the downstream system now have two records, or one? Every action with a side effect needs an idempotency key that the agent controls and the tool honours. Without it, the retry logic you added for reliability is a duplication bug you added for reliability.

## 2\. What happens on the twentieth attempt?

Retries need a ceiling, backoff, and somewhere to go when they are exhausted. A run that fails permanently should land in a queue a human can see, not in a log line. The question "how would anyone find out this failed" should have an answer that is not "a customer tells us".

## 3\. How does it escalate?

Escalation is a feature, not a failure. It needs a defined trigger, a destination, and enough context attached that the person receiving it does not have to reconstruct the run. An agent that cannot hand off is an agent that will guess, and guessing is how the incidents happen.

## 4\. Can you reconstruct any single run?

Six weeks later, someone asks why a particular invoice was routed for approval. You need the inputs it saw, the tool calls it made with resolved arguments, the outputs it produced, the model and prompt version, and the timing. If any of those are missing, the answer is a shrug, and a shrug is expensive when the question comes from a regulator or a customer's lawyer.

## 5\. Does a change have to pass anything?

Prompt edits are code changes with none of the safeguards. If a prompt can reach production without passing an evaluation suite, then the system's behaviour is protected by whoever is careful that day. The suite should run in CI, gate the merge, and report per-segment rather than in aggregate.

## 6\. What does it cost, and what stops it?

Per run, at the median and at the 99th percentile. Then: what is the ceiling, where is it enforced, and what happens when it is hit. An agent with a loop and no budget cap is a bill with no upper bound, and the loop is usually discovered by the bill.

## 7\. What happens under load?

Provider rate limits, concurrency limits on your side, and queue depth. The interesting question is what the system does when it is saturated — degrade, queue, or fail — and whether that behaviour was chosen or inherited.

## 8\. What data does it touch, and where does it go?

Which fields reach the model provider. Which are redacted before they do. What retention applies to prompts, completions, traces, and any derived artefacts — and whether those retention periods are actually different, because they usually should be. This has to be written down before launch, not reconstructed during a security review.

## 9\. Can you turn it off?

A flag that disables the agent without a deploy, and a defined fallback for what happens to work in flight. "We would revert the commit" is not a rollback plan when the queue has four thousand items in it.

## 10\. Who owns it on a Tuesday?

A named team, an alert that reaches them, and a runbook that describes the three most likely failures and what to do about each. Systems without an owner do not get maintained; they get replaced, eighteen months later, at full cost.

## 11\. How will you know it got worse?

Not "did it break" — got worse. Quality drifts without erroring. You need a metric that would move: sampled accuracy on recent production traffic, escalation rate, refusal rate, or a downstream business measure. Something with a threshold and an alert, checked on a schedule, so that a slow decline is visible before it is severe.

## Using the list

Most teams we talk to have answered items 1, 4 and 10 and none of the others. That is not a criticism — the remaining eight are invisible until they are urgent, which is precisely why they are worth going through deliberately while nothing is on fire.

The list is also a reasonable scoping tool. If a project has budget for the model work and not for these, it does not have budget to put the model work into production, and it is better to find that out in the first conversation.