In From Cocoa to Chocolate, the two gateways govern the models agents think with and the tools they act on. The Agent Orchestrator is the production line itself: where agents are built, run and kept on the record. This piece walks through it in full.
A workforce of specialists, not one chatbot
A generalist model knows everything about the internet and nothing about the business it is meant to serve. Asking one generic chatbot to do everything means asking it to know every policy, every system and every line it must not cross, all at once.
The future of enterprise AI is not that chatbot. It is a managed workforce of specialists, each tuned to a distinct job, grounded in the knowledge that job depends on, and given only the tools it needs. The Agent Orchestrator is where that workforce is built, organised and governed.
The Orchestrator builds and runs its own agents. It does not take charge of agents built elsewhere; apps and agents from other vendors come under control through the AI Gateway.
Everything an agent needs, in one place
Each agent is assembled from the same set of parts, configured side by side. Because every agent is built the same way, every agent can be inspected, compared and corrected the same way.
Instructions
A system prompt that sets the agent’s role and rules.
A model
Chosen per agent, from the models its project has turned on.
Knowledge
The knowledge bases and documents the job depends on, so answers come with a source instead of a guess.
Tools
Built-in capabilities the agent can use.
MCP connections
The systems it can reach, with the actions it is allowed to take.
Integrations
Links to the other apps the business uses.
Memory
Whether it remembers earlier conversations.
Schedule
Jobs it runs on its own, at set times.
Collaboration
The other agents it works with.
This is the assembly stage of the factory. When something goes wrong later, the team does not rebuild the agent. It finds the one part responsible and changes that.
Limits on how far it runs
Beyond the prompt, each agent’s behaviour is tuned with model parameters: temperature, maximum tokens, the maximum number of step iterations, and a limit on how many tool steps it may take in a row.
An agent caught in a loop keeps calling, and keeps billing, until someone notices. With run limits, an agent that would otherwise loop stops where the business told it to.
An identity for every agent
Each agent gets its own key, so every model call, every tool call and every run is tied to the agent that made it. Nothing an agent does is anonymous.
That is what makes the rest of the factory work. Spend can be traced to the agent that caused it, access can be granted to the agent that needs it, and a run can be read back knowing exactly who ran it.
Where people work
Agents meet people where they already are: in Microsoft Teams, in WhatsApp, inside the business’s own apps through the Run API, or on a schedule with no one asking. Nobody has to learn a new place to work to use them. Wherever an agent is reached, it is the same agent, with the same instructions, limits and identity.
Agents that work together
One agent can hand a task to another, then carry on with the result. A business can build specialists that each do one thing well, and let them cooperate, rather than overloading one agent with every job.
When the steps must not vary, an agent can also hand off to a fixed process built in the Business Process Studio.
Every step of every run, on the record
Any run can be opened and read in order.
- What was asked
- Each step, with its timing
- The tool called and what it sent
- The system it reached
- What came back
- The final answer
The final answer can look convincing even when the work behind it was not. When an answer is wrong, the step where it went wrong is visible, so agents are tuned with evidence instead of guesswork. The full monitoring view is covered in Seeing What Every Agent Did, and Why.
The harness
The two gateways wrap the Orchestrator: one for the models its agents use, one for the tools they call. The Orchestrator’s agents inherit the gateways’ rules directly.
The harness
- Model side, through the AI GatewayManaging model providers, fallback to another model or host, virtual keys for teams and members, budgets per team, member or key, keys that expire, and spend caps with a circuit breaker.
- Tool side, through the MCP GatewayThe tool registry, authentication, tool filtering, access control such as allow read and deny write, metering and budgets, and an audit log.
The Orchestrator governs the agents it builds and runs, and the harness routes those agents through both gateways.
What it changes
Business teams shape their own AI. The people who understand the work write the instructions, choose the knowledge, decide which steps need a person, and correct the agent when it goes wrong. They build specialists for their own jobs rather than waiting in line for a generic assistant.
In chocolate terms, the chocolatiers write their own recipes. The full feature set is on the Agent Orchestrator page.
The short version
- Every agent is assembled from the same parts, so every agent can be inspected and corrected the same way.
- Run limits stop loops, and each agent’s own key means nothing it does is anonymous.
- Agents work together, meet people where they work, and leave a full record of every run.
- The harness routes the Orchestrator’s agents through both gateways.



