Infinigence AI and "Three-Body" Release Sophon Agent 1.0 for AI Characters

Infinigence AI (无问芯穹), a Chinese company that builds infrastructure for AI computing, and the rights holder of the Three-Body universe (三体宇宙) have jointly released Sophon Agent 1.0 (智子Agent). The product is positioned as a platform for character agents based on the science-fiction IP. Its foundation is what the authors call Agentic Infra: infrastructure tuned for running agents rather than for one-off chatbot replies.

The name is no accident. In Liu Cixin's novel, the sophon (智子) is a supercomputer folded into a proton, and one of the central symbols of the series. Using it as a brand for an AI platform is a clever move: readers immediately make the intended association.

Where character agents break

Over the past year, Chinese online platforms have rushed out interactive products built on role-play agents: dialogues with characters from films and TV series, virtual artists, and interactive dramas with AI participants. The market looks lively, but most of these products suffer from the same problems. The release authors name three main ones:

• weak character fidelity: the character does not speak the way they would in canon;
• persona drift: after several rounds of dialogue, replies move away from the assigned role, and the character turns into a faceless assistant;
• degradation under load: when many people talk to the agent at the same time, lag appears and the rhythm of the conversation breaks.

The first two points concern the quality of role modeling; the third concerns serving engineering. They are usually treated separately: with prompts and fine-tuning on one side, and cluster scaling on the other. Sophon Agent bets on solving both problems at the level of a single infrastructure.

Why the problem comes down to infrastructure

For a role-play agent, context is the main resource. For the character to remember their biography, manner of speech, relationships with other characters and plot events, all of it has to be fed to the model at every step. Long context, memory and request routing therefore become the core of the product, not a secondary detail. The more users and the longer the dialogues, the faster the inference load grows, and the more noticeable the delays become.

It is therefore logical that the release comes not from a model developer or a studio, but from a company in the computing infrastructure world. Infinigence AI specializes in optimizing and orchestrating compute, and for it Sophon Agent is a showcase case: proof that the infrastructure layer can be responsible for user-facing quality, not only for the cost of a token.

The calculation is clear for the IP owner too. "Three-Body" is one of the most recognizable Chinese science-fiction brands, with a ready-made audience and a strict canon. An agent that speaks in a character's voice and does not break character turns the universe from an object of consumption into an interactive environment. At the same time, any OOC (out-of-character) line hurts the brand more than it would with an ordinary chatbot, so the requirements for persona control are higher here.

What teams building characters should watch

Teams building their own character agents can take a practical checklist from this release. Measure character fidelity not on a single reply but on long dialogues: drift shows up in the fifth to tenth round, not the first. Memory and the canonical knowledge base about the character should live outside the prompt and be pulled in on request, otherwise the context bloats. Load testing should use real scenarios with hundreds of parallel sessions, because latency in a live dialogue breaks the sense of presence more than an inaccurate detail does. And finally, agree on the limits of acceptable character behavior with the rights holder in advance: infrastructure can enforce boundaries, but a human has to set them.