Agents Added One at a Time
AI agents tend to arrive one at a time. One pulls data out of documents, another drafts replies, a third reaches internal data. Each is built by different people and wired to the model provider on its own.
Then the questions pile up. Users ask which bot they should talk to. Every agent carries its own provider key and its own idea of who may use it. The model bill arrives once a month and says nothing about who spent what. Adding one more agent means new wiring, new keys and new screens.
Three Parts and a Gateway
The platform consists of three parts, plus one gateway that every model call passes through.
- Hub. It signs people in and knows each person's role, message limits and the agents they may use. Sign-in works by password or through Google and Microsoft.
- Orchestrator. It runs the conversation, decides from the context which agents to call, calls them and assembles the answer.
- Agents. Each does its own job and speaks the Model Context Protocol, so an agent built elsewhere connects the same way as one built in-house. An agent joins the catalog by being deployed with a label, without anyone editing a list.
- Gateway. Every call to a model, from the orchestrator or from an agent, leaves through one gateway. It holds the only provider key and counts what goes out.
The hub checks every agent in the background and stops offering any agent that fails to answer. While an answer is built, a status line names the current step in words. Above the answer, the person sees which agents it came from, the time and tokens of each step, and, one click deeper, the exact question sent and the reply.
An administrator sets access per person with two switches per agent: whether the orchestrator may use it for them, and whether a direct dialog is open. Agent credentials can be written but never read back.
Shortcuts That Stop Working at Scale
- One bot with every tool attached. The prompt grows with each tool, every tool is offered to every user, and there is nowhere to say who may use what.
- Access written into the prompt. A sentence in a prompt asks the model to avoid something. It does not stop it.
- Spend taken from what agents report. An agent that spends and stays silent is invisible without an independent count.
- A list of agents kept by hand. The list drifts from what is actually running, and a dead agent is still offered to users.
- A spinner while several agents work. The person cannot tell whether the system is thinking, waiting on an agent or stuck.
Technology Stack
Access Decided on the Server
The prompt guides. The server forbids. Instructions, rules and skills shape an answer, but they forbid nothing. Which agents the orchestrator may call for a person, how many messages they have and who is an administrator are checked by the server on every request.
- One exit to the models. The provider key exists in one service only. The orchestrator and the agents hold keys to the gateway, never the provider's.
- Spend is counted twice. The platform counts per person and per period. The gateway counts everything that goes to the models on its own, and a gap between the two totals shows up as a warning.
- Closed by default. A service with a missing key or secret refuses to start, and an agent without its trust secret rejects every call.
Platform Guarantees
This is Alltegrio's own product, and the points below are its design specification.
- No model call that bypasses the gateway.
- Total spend measured at the gateway, separately from the platform's own count.
- A new agent added by deploying it with a label, with no change to the hub.
- An agent that stopped answering is not offered to conversations.
- A change of rules or instructions applies from the next answer, without a restart.
When a Shared Runtime Pays Off
A shared runtime starts to pay for itself once agents stop being one experiment and become several services people rely on:
- several agents built by different people, each wired to the provider on its own;
- a monthly model bill that says nothing about who spent it;
- users who ask which bot they should be talking to;
- agents meant for customers as well as staff.
Service companies, support centers, back offices and software vendors usually reach that point first. An agent that already speaks MCP can be pointed at the gateway and deployed with a label, without a rewrite. For a client, the same design can be set up on their own cloud, model providers and sign-in. Read more about agent orchestration and the Model Context Protocol in our glossary, or about how we deliver this on our AI agent development and LLMOps pages.
exit to the models, designed for every call from every agent
independent counts of model spend, by the platform and by the gateway
access switches per agent and person: orchestrator use and direct dialog