Building New Agents

Register inputs for a new agent and let Maniac bootstrap completions and build a model.

Most Maniac workflows start with an existing LLM call — you capture production traffic, then optimize. But what if you're building something new? You know the kinds of requests your agent will handle, but you don't have completions yet.

Maniac supports input-only registration. You register the inputs your agent needs to handle — without any outputs — and Maniac's infrastructure generates completions, bootstraps a training dataset, and builds an optimized model for you.

This is especially useful for high-throughput agents where you can enumerate or sample the input space upfront.

---

How It Works

1. Create a container with a base model and system prompt that describe the agent you want to build.

2. Register inputs — just the messages, no output field needed.

3. Maniac handles the rest — it generates completions, bootstraps a training dataset, and automatically optimizes a model for your task. No manual steps required.

---

Step 1 — Create a Container

Set the initial_model to the frontier model you want Maniac to bootstrap from, and the default_system_prompt to the instructions for your new agent.

Python

from maniac import Maniac

maniac = Maniac()

container = maniac.containers.create(
    label="ticket-router",
    initial_model="openai/gpt-4o",
    default_system_prompt=(
        "You are a support ticket router. Given a customer message, "
        "respond with exactly one label: billing, technical, account, or general."
    ),
)

TypeScript

---

Step 2 — Register Inputs Only

Each item only needs an input — the output field is optional. When omitted, Maniac treats the entry as an input that needs a completion generated.

Python

inputs = [
    "I was charged twice for my last order",
    "The app crashes when I try to upload a file",
    "How do I change my email address?",
    "Do you offer student discounts?",
]

maniac.chat.completions.register(
    container="ticket-router",
    items=[
        {
            "input": {
                "messages": [{"role": "user", "content": text}]
            }
        }
        for text in inputs
    ],
)

TypeScript

That's it — no output field, no completions to generate yourself.

---

Uploading Inputs at Scale

If you have a large corpus of representative inputs (e.g. historical support tickets, logged queries, synthetic samples), upload them in batches to avoid timeouts.

Python

BATCH_SIZE = 1500

for i in range(0, len(all_inputs), BATCH_SIZE):
    batch = all_inputs[i : i + BATCH_SIZE]

    maniac.chat.completions.register(
        container="ticket-router",
        items=[
            {
                "input": {
                    "messages": [{"role": "user", "content": text}]
                }
            }
            for text in batch
        ],
    )

    print(f"Uploaded {i + len(batch)} / {len(all_inputs)}")

TypeScript

---

What Happens Next

Once your inputs are registered, add an Eval from the Evals tab inside your container to define how the agent's outputs should be judged (judge prompt or code eval). From there, Maniac takes over automatically — it generates completions from your base model, builds a training dataset, and optimizes a new model in the background. When a model is ready, it will appear in the Models tab inside your container, where you can deploy it so inference requests are routed through it.

The more representative your inputs are of real production traffic, the better the resulting model will perform. If you can, sample inputs that cover the full range of scenarios your agent will encounter.