FAQ
What teams ask before an agent goes live.
Agency is how you build and govern agents on Czar. Substrate is the platform underneath. Consulting is the practice, when you want engineers who already know that stack.
- What is an agent on Agency?
- A versioned sequence of work. It takes an input your desk already understands, reads what it is allowed to read, drafts or decides, waits where a person has to own the outcome, and calls only the systems you described. Publishing freezes that sequence. A deploy points a stable address at one published version.
- Where does the work actually run?
- On Czar, the appliance. Agency is the control plane: compose, publish, deploy, approve, and read the run. The model inference and the material stay on the appliance. If Czar is air-gapped, the agent is air-gapped. If you have allowed the appliance to reach specific secured systems, the agent can use the ones you named.
- Which models can it use?
- The models you host on the appliance. An admin chooses which of those a given agent may call. If that model is not being served, the run reports it and stops. Agency does not send the step to a model vendor.
- How is this different from Substrate?
- Substrate is the platform on the appliance: grounded answers, memory, models you host, and a record of what a person asked. Agency is how you turn that into an agent that runs a job — with versions, tools, a person in the path, and a run record. They sit on the same appliance.
- How does an agent reach our other systems?
- You describe the call and publish it with the version: where it goes, what it sends, which credential it may use. That list is what the run can touch. Credentials are usable by the run and are not readable back out. Adding a system is a new description in a new version.
- Who has to approve what?
- You decide, on the design. A step can require a role before the run continues. Policy can also allow, flag, or stop a step without a person. If a step must be checked and the policy service cannot be reached, that step waits.
- Can we run more than one agent?
- Yes. Each agent has its own design, its own published versions, and its own address. Each address has one live version. Applications call the address, and you choose when that address moves to a newer publish or back to an older one.
- What if we publish a bad version?
- You point the address back at an earlier published version. The previous design is still there, because publishing kept it. Drafts are where you fix the next version. Live traffic keeps calling a version you already froze.
- What does someone see when they open a run?
- The sequence. Input, each step, model output, passages used, tool calls, the policy decision, and the person who approved when a person had to. They can follow it live or open it after it has finished.
- Can our own people build this?
- Yes. Operators compose and publish. Developers can keep the same agent as a file. Reviewers sit on the steps that wait. Admins decide models, who may publish, and which tools exist. E31 Consulting can also put certified engineers on the use case — with your team, or as a project handed back running.