What Makes an AI Agent "Production"?
A production AI agent handles real conversations with real consequences, so it needs limits on what it accepts, checks for abuse, rules for what it may say, logs of every model call, a way to switch it off and a clear hand-off to a person. The model is the smallest part of the work.
This post uses our own sales agent as the example, because it is the production agent we can show in full. Every capability below is something we checked in our code on 2 October 2026.
A demo chatbot versus a production agent
Our view, from running one
- Accepts any input of any length
- Answers everything, including what it does not know
- No record of what it said or what it cost
- Limits input, rate and abuse before the model sees it
- Follows written rules and hands off when it should
- Logs every model call and can be switched off
What Does Our Own Sales Agent Do?
Vector is the AI sales agent on pixelmojo.io. It answers visitors in the chat widget and replies to inbound email from the same core, qualifies leads across 12 dimensions, adapts to emotional cues, and hands serious or unhappy conversations to our founders.
Vector, in numbers
pixelmojo.io codebase, 2 October 2026
qualification dimensions
documented negotiation patterns
channels from one core: chat and email
provider hub for every model call
The 11 negotiation patterns are written guidance on how to handle common moments, such as a direct pricing question or a visitor who says the price is too high. They shape tone and order, not facts: prices and terms come from the agent's rules, which set out our current offers.
Which Business Problems Does It Solve?
An AI sales agent solves the problems of first contact. It answers the first questions at any hour, asks the qualifying questions a salesperson would, explains prices and terms the same way every time, filters out spam, and routes serious leads to people. It does not replace the people; it decides who they should talk to next.
| Problem | What the agent does | What still needs a person |
|---|---|---|
| Visitors arrive outside office hours | Answers immediately, in chat or by email | Anything that needs a decision or a contract |
| Leads vary widely in fit | Asks qualifying questions across 12 dimensions | The final judgment on fit |
| Prices and terms get explained inconsistently | States current offers from its rules | Custom scope and negotiation |
| Spam and throwaway inquiries waste time | Flags spam patterns in chat and recognizes disposable email addresses | Edge cases it flags |
| Unhappy visitors need care | Tries to de-escalate, then hands off if anger continues | The conversation after the hand-off |
We do not publish outcome numbers for Vector here. An earlier version of this post opened with an unsourced claim that sales teams waste 60% of their time on unqualified leads; we removed it.
What Guardrails Does a Production Agent Need?
A production agent needs guardrails at every step: before the model sees the message, around the model call, and in what the agent is allowed to say. These are the ones Vector runs today.
The path of one message
How Vector handles a request, October 2026
Limits
length, history and rate
Checks
spam and email quality
Model
gpt-4o-mini via the hub
Rules
what it may say
Hand-off
to a person when needed
| Guardrail | What it does in Vector |
|---|---|
| Input limits | Rejects messages over 2,000 characters and keeps only the last 12 messages of history |
| Rate limits | Caps messages per minute and per day |
| Spam detection | In chat, flags spam patterns before a reply is generated |
| Email quality | Recognizes disposable and free-provider email addresses |
| Provider hub | Every model call goes through one file that logs it; the agent checks a kill switch first |
| Conversation cap | Ends a chat after 30 messages with a pointer to email the founders |
| Rules | Never says clients own the AI models, vector database or training data, which they do not; hands questions it cannot answer to Lloyd |
| Emotional hand-off | Tries to de-escalate anger, and hands off to a person after two attempts or when low energy persists |
The provider hub matters most. Because the lint configuration blocks provider SDK imports and the known provider API addresses everywhere except that file, new code cannot add those direct paths to a model by accident. We describe that approach in Thread-Based Engineering.
Does Retrieval Stop Hallucination?
No. Retrieval-augmented generation reduces hallucination by giving the model relevant passages to work from, but it does not stop it. A preregistered 2024 study of commercial legal research tools that use retrieval found they still hallucinated a substantial share of the time, even though they did better than a general-purpose chatbot.
The study is on arXiv. Our own use matches its lesson. Vector does not use retrieval today; it carries its knowledge in its instructions. Retrieval powers a different feature on our site, the Ask search over our blog, which embeds a question and finds matching passages in a Supabase pgvector table. In both cases the rules and the sources do the work of keeping answers honest, not the retrieval alone.
Where Do Multi-Agent Systems Fit?
Multi-agent systems fit work that splits into specialist steps that have to coordinate, such as an intake step, a pricing step and an escalation step. Hive is our multi-agent platform, which we run and also offer as a 12-week client build. The insurance and logistics agent teams on the Hive page are illustrative configurations of a client build, not agents we operate for those industries. Our guide to multi-agent systems covers when one agent is not enough.
Should You Build or Buy an AI Sales Agent?
Buy when an off-the-shelf tool covers your questions and your CRM. Build when the agent has to know your offers, follow your rules and hand off the way your team works. As a reference point, we build Vector for clients as a 60-day project: $35,000, plus $1,500 a month for the intelligence platform with a 12-month minimum, and 5% of deal value from Vector-qualified leads that close. The client owns the application code. Our post on how we built the multi-channel sales agent covers the build itself.
What Changed in This Post?
We rewrote this post on 2 October 2026 as part of a review of every post on the blog. The changes:
- it said our agent ran GPT-4o; it runs gpt-4o-mini;
- it said retrieval eliminates hallucination; it reduces it, and the post now cites a study that measured how much remains;
- it described insurance-quote and logistics-tracking agents processing real business in production; those are illustrative configurations, so the claims are withdrawn;
- it said retrieval was core to the sales agent; the agent does not use retrieval today;
- an unsourced statistic about sales teams' time and an unsourced claim about leads Vector eliminates are removed.
Production AI Agents: Questions Readers Ask
Common questions about this topic, answered.
What makes an AI agent production-ready?
A production agent handles real conversations with real consequences, so it needs limits on what it accepts, checks for abuse, rules for what it may say, logs of every model call, a way to switch it off and a clear hand-off to a person. The model is the smallest part of the work.
What is Vector?
Vector is the AI sales agent on pixelmojo.io. It answers visitors in the chat widget and replies to inbound email from the same core, qualifies leads across 12 dimensions, adapts to emotional cues and hands serious or unhappy conversations to our founders. We also build Vector for clients as a 60-day project.
Which model does Vector run on?
Vector runs on OpenAI gpt-4o-mini, called through a single provider hub in our codebase. The hub logs every call, the agent checks a kill switch before it calls the model, and lint rules stop other code from calling a model provider directly.
Does retrieval-augmented generation (RAG) eliminate hallucination?
No. It reduces it. A preregistered 2024 study of commercial legal research tools built on retrieval (arXiv 2405.20362) found they still hallucinated between 17% and 33% of the time, though less than a general-purpose chatbot. Retrieval needs rules, sources and review on top of it.
Does Vector use RAG?
Not today. Vector carries its knowledge in its instructions: our services, prices, terms and the rules for what it may say. Retrieval powers a different feature on our site, the Ask search over our blog, which embeds questions and finds matching passages in a Supabase pgvector table.
What does an AI sales agent do for a business?
It answers first questions at any hour, asks the qualifying questions a salesperson would, explains prices and terms consistently, filters spam and routes serious leads to people. It does not replace the people; it decides who they should talk to next.
Does Hive run insurance and logistics agents in production?
No. Hive is a running multi-agent platform and a 12-week client build. The insurance and logistics agent teams on the Hive page are illustrative configurations of what a client build could look like, not agents we operate for those industries.
What changed in this post on 2 October 2026?
We rewrote it. It said Vector ran GPT-4o (it runs gpt-4o-mini), that retrieval eliminates hallucination, and that our platform processed insurance quotes and logistics shipments in production; all three are corrected or withdrawn, along with a statistic about sales teams that had no source.
