# Agent Orchestrator — production AI agents in Open Mercato

Source: https://www.openmercato.com/agent-orchestrator

# Agent Orchestrator - the new AI agent module

Everyone has an AI agent demo. Almost nobody has one you can safely let into a real business — real data, real processes, real customer money. A demo runs in a sandbox; production runs on your permissions, audit trail, and liability. Agent Orchestrator closes that gap — out of the box with Open Mercato Enterprise.

## Run your agent where you already run

No walled garden. Pick the runtime that fits your stack. Your agent, your environment. We adapt to it — not the other way around.

### OpenCode

Runs inside your Open Mercato Docker container.

### In-process

Lightweight agents, no extra infrastructure.

### Cloud runtimes

Azure AI Foundry, Google, Amazon and others via one connector.

## An agent becomes a function in your workflow

The Agent Orchestrator plugs into Open Mercato Workflows through a new node — Invoke Agent. No glue code, no separate integration project.

### Pass context

The workflow hands the agent a well-defined context.

### It does the work

The agent runs as any other function.

### Workflow waits

It resumes on the agent's response.

## Two controls make it production-grade

### Confidence thresholds

Set the bar. Above your threshold the agent decides and the workflow moves on. Below it, the decision escalates to a person. You decide how much autonomy you hand over.

### Human-in-the-loop

When output needs sign-off, the workflow stops and routes it to the right person. The agent proposes, the human disposes.

> Runs on top of the Open Mercato MCP Server — the tools you already expose are discoverable and callable by agents. No new tooling to wire up.

## Trust, because you can see everything

Most platforms ask you to take their word for it. We log instead. Every agent action is audited exactly like a human action — a data change by an agent is traced the same way as one made by a person. Nothing off the record.

- 01Autonomous action - a clean agent-level record.

- 02Person-triggered action - both the agent and the person who set it in motion.

You always know what the agent did — and who put it to work. For regulated buyers, that's the price of entry.

## Stop guessing whether your agent is any good

### Agent Evals

You don't ship code without tests. Define evaluation scenarios, benchmark your agents against them, and watch how they hold up on real cases — in the same place they run.

### Observability

See which model an agent used, what it cost in tokens, and how it performs over time. Admins and AgentOps teams tune behavior with real numbers.

## How is this different from n8n?

Fair question — people ask it first.

The difference between automating a step and trusting an actor.

|  | In n8n | In Open Mercato |
| --- | --- | --- |
| What an agent is | Another node that calls an API | svg]:block [&>svg]:size-full size-5 [&_[stroke-width]]:[stroke-width:var(--icon-stroke)] [&>svg]:[stroke-width:var(--icon-stroke)]" style="--icon-stroke:1.68"> A supervised participant in the process |
| How it runs | Fires a request, hopes for the best | svg]:block [&>svg]:size-full size-5 [&_[stroke-width]]:[stroke-width:var(--icon-stroke)] [&>svg]:[stroke-width:var(--icon-stroke)]" style="--icon-stroke:1.68"> Runs inside your data, permissions and audit |
| Relationship | Bolted on, outside your system | svg]:block [&>svg]:size-full size-5 [&_[stroke-width]]:[stroke-width:var(--icon-stroke)] [&>svg]:[stroke-width:var(--icon-stroke)]" style="--icon-stroke:1.68"> Governed, native to the platform |

## One layer, not three features

Runtimes, workflow integration, audit, evals — not four separate things. One coherent layer, start to finish.

- 01CreateDefine an agent, its tools and the runtime it runs on.

- 02RunAs a standalone function or as a step inside a workflow.

- 03TraceEvery action audited, two layers deep — agent and person.

- 04UnderstandModel, token cost and performance, all in one place.

- 05BenchmarkRun evals against real-world scenarios.

AI agents spent two years doing impressive demos. This is the boring, hard part nobody screenshots — getting them safely into production.

## Software is about to be built completely differently.