AutoGen with unoblox (India)

AutoGen builds applications out of multiple agents that converse with each other to reach a result. Every message between agents is a model call, so a multi-turn AutoGen conversation can involve far more requests than a single chat exchange. Pointing that traffic at unoblox keeps all of it on one rupee-denominated account regardless of which model each agent uses.

How AutoGen's config_list maps onto unoblox

AutoGen agents take a config_list describing which model, base URL, and key to use. Because unoblox is OpenAI-compatible, this is the same shape AutoGen already expects for any custom endpoint:

config_list = [
    {
        "model": "openai/gpt-5-mini",
        "base_url": "https://api.unoblox.ai/v1",
        "api_key": "ub-gw-xxxxxxxxxxxxxxxx",
    }
]

assistant = AssistantAgent(name="assistant", llm_config={"config_list": config_list})

A second agent, say a critic or a user-proxy, can point at a different config_list entry with a different model id, so the two sides of the conversation do not have to run on the same tier.

Assigning models by conversational role

Agent role	Model	₹ per 1M tokens (input / output)
High-volume assistant turns	DeepSeek V4 Flash	₹9.07 / ₹18.14
Balanced assistant	GPT-5 mini	₹25.2 / ₹201.6
Critic or verifier agent	Claude Sonnet	₹201.6 / ₹1008
Free agent for scripted tests	Qwen3 1.7B	₹0

Anything outside this table, see live ₹ pricing on /models before wiring it into a conversation that runs automatically.

Controlling how long a conversation runs

AutoGen conversations can loop between agents until a stopping condition is met, and each loop iteration is billable. Set a max-consecutive-auto-reply limit, or your framework version's equivalent, on every agent before you let a multi-agent conversation run unattended, and log which agent and model produced each turn so a runaway back-and-forth is visible in your own logs before it shows up as a surprise on the monthly invoice.

Testing a new agent setup cheaply

Point every agent in a brand-new AutoGen graph at the free model first, confirm the conversation terminates correctly and produces the shape of output you expect, then upgrade only the agents whose output quality actually needs it. This catches infinite-loop or malformed-prompt bugs before they cost anything.

Frequently asked questions

Does AutoGen need a plugin to work with unoblox? No — AutoGen's config_list already supports any OpenAI-compatible base_url and api_key, which is exactly what unoblox provides.

Can I run a fully open-weight agent conversation? Yes — models such as Qwen3.8-27B are open-weight and available through the same endpoint, alongside proprietary models, if you want every agent on open-weight models.

Is our conversation data kept in India? Only for calls routed to the small unoblox-hosted models; external models like GPT and Claude process the conversation on their own infrastructure, with unoblox providing rupee billing and a GST invoice rather than India data residency.

How do we avoid two agents looping forever and running up cost? Set explicit turn limits and termination conditions in AutoGen itself; the gateway bills every call it receives, so cost control has to happen in your agent configuration.

Can each agent in the graph use a different model provider? Yes, each config_list entry is independent, so you can mix GPT, Claude, DeepSeek, and Qwen agents in the same conversation graph.

Get started in rupees → https://unoblox.ai/sign-in