AI agents · Orchestration · Production notes

AI agents and orchestration: what I run in production

A useful business AI agent is a language model wired to tools, data and guardrails, doing a task without anyone watching. Here are the four systems I run: what they do, what they run on, what stops them from making things up, and what still breaks.

Published 2026-10-01. Client projects go through ZAOURI, my agency.

zaouri.com's sales agent

Answers visitors in French, English and Arabic, qualifies the need, then saves a lead brief with the save_qualified_brief tool.

What it runs on: gpt-oss-120b on Cerebras' free tier, falling back to Groq (same model family, so one prompt to tune), then a WhatsApp card if both quotas are exhausted: the visitor never sees an error.
Guardrail: Every amount, duration or percentage the model writes is checked against the published rate card before display; a figure that is not on the card is masked. A brief is saved only if the server validates it field by field: the model cannot claim a field it did not collect.
Status, dated: Live since July 2026. What breaks: this model class fires the tool too early, routinely; the server rejects the incomplete brief and the conversation continues.

Pulse: the SEO collector that runs on its own

Every morning it queries the Search Console API for the sites I operate, stores the rows in SQLite, compares two windows and writes alerts. On Mondays, one report per site.

What it runs on: Python, launchd at 9:30, no human or AI session in the loop. A citation probe is plugged in: Gemini with Google Search grounding, on a frozen set of questions (20 for this site, 22 for ZAOURI).
Guardrail: "Cited" means "listed in the answer's sources", not "named in the text". Several runs per question and a Wilson interval, because an engine never returns the same sources twice.
Status, dated: Collection running since late August 2026, citation probe since 2026-10-01. Stated limit: an API does not answer like the interface the client sees (see below).

The seo-constellation MCP server

Gives Claude 16 tools: Search Console (5), GA4 (4), on-page and schema audit (2), reads of the pulse database (5). I ask in plain language, the assistant calls the real data.

What it runs on: Python, Model Context Protocol, read-only Google service account.
Guardrail: Nothing is scraped: official authenticated APIs only, so a third-party tool breaking cannot pass for an algorithm update in my numbers.
Status, dated: Used for every audit and every check on this site.

Marrakech Private's machine

7 scheduled workflows for marrakechprivate.com: listing freshness, event watch, queries, demand, AI citations, prospecting, competitors.

What it runs on: One daily launchd job; for each workflow due, scripts prepare the data, then Claude (Haiku) runs a versioned prompt with an allow-list of tools; every run is logged.
Guardrail: Tools are restricted per workflow (an audit can only write through its validation script). A failing workflow never blocks the next one.
Status, dated: The log shows successes and failures: 15 venues checked on 2026-09-21 and 2026-09-28, 87 events triaged on 2026-09-29; other runs still stop on their turn limit.

What breaks, and what it changed

An agent is judged by its failures as much as its successes. The four that keep coming back:

The tool fired too early

A model triggers "save the brief" before it has the answers. Fix: server-side validation, never the schema alone.

Free-tier quotas

No free tier can carry a sales widget alone: a chain of providers, then an exit that never shows an error.

The turn limit

A scheduled agent that runs out of turns stops midway. Fix: prepare the data with scripts before the call, and log the failure.

The API is not the interface

Measured overlap between an engine’s API answer and its interface answer: 12 to 14.8% (arXiv 2609.18729). An API probe gives a series comparable over time, not what the client sees.

Source for the API / interface figure: arXiv 2609.18729 (v1, 2026-09-16, 2,528 queries, correlational). The full measurement method: AI citation tracking.

Frequently asked questions

What is a business AI agent?

A language model wired to tools (a database, an API, a form) and to rules, doing a task end to end: qualifying a lead, collecting data every day, checking listings. Unlike a chatbot, it acts, and what it does is checked server-side.

Which model do you use?

Whichever fits the task and the budget. zaouri.com's agent runs on gpt-oss-120b through Cerebras' then Groq's free tiers; the scheduled workflows use Claude (Haiku for repetitive tasks). I do not train models: I integrate existing ones.

Can an agent make up a price or a promise?

Yes, that is a sales agent's first risk. On zaouri.com, every amount, duration or percentage the model writes is checked against the published rate card before display, and anything not on the card is masked.

What is an MCP server?

The Model Context Protocol lets an assistant like Claude call tools. My seo-constellation server exposes 16 of them (Search Console, GA4, audit, pulse database), through official APIs only.

Do your agents really run unattended?

Some do: pulse has collected every morning since late August 2026. Others still fail, and their log shows it next to the successes. An agent without a failure log is an agent whose breakdowns you never see.

How much does a custom AI agent cost?

Projects go through ZAOURI, my agency. The cost depends on the tools to connect and the volume; it is priced after scoping, not before.

An agent for your business?

Projects are scoped and priced at ZAOURI, my agency. GEO and the method stay documented here.

See ZAOURI services The GEO-360 method