Apollo is an agentic Language Model that replaces the transformer with a Neuro-symbolic architecture built for agents. Apollo enables conversational agents that be fine-tuned i.e. can learn and evolve from human feedback, resulting in ever-improving performance. Developed over the past six years in collaboration with 60,000 human agents, Apollo outclasses traditional LLMs across agentic use-cases. Apollo is trained on a neuro-symbolic language that constitutes both descriptive and procedural (“agentic”) data, replacing autoregressive inference with neuro-symbolic reasoning, which is better suited for agents. This method relies on obtaining a structured interaction state, achieved through sensory data collected to produce a symbolic, parameterized representation of each interaction.
Apollo enables companies of any kind to deploy fine-tuned agents, versions of Apollo that are fine-tuned on a specific task, and can continuously improve with human feedback. They further feature fine-grained control, the capacity to adhere to the deploying company’s policies and a white-box view of their decision-making and reasoning. They offer superior tool use, steerability and overall performance over LLM Agents.
Apollo does not train on company data, instead, it can be pre-integrated with any number of internal systems, and call the relevant API endpoint in every message. Like transformer-based LLMs, Apollo has a “system prompt”, broken down to structured, conditional policies, controlled via the “Instructions” tab. Tools (“Workflows”) are pre-inserted to Instructions and only visible to the model when necessary to complete a specific interaction successfully. This enables 100% success rate in executing complex API calls and the ability to return grounded answers in each interaction via functioning tool use over many different endpoints.
Unlike transformer-based LLMs, Apollo is fine-tuned via human feedback at the interaction level. In the Simulator, companies can generate automated conversations simulating the agent in production in front of users. They then proceed to fine-tuning the agent by correcting its actions and behavior in each unsatisfactory interaction. This process can be repeated continuously to improve the performance and alignment of the agent.
Unlocked with Apollo are some exciting new agentic use-cases, including the ability for companies to responsibly communicate with their customers using conversational AI, and to sell their products and services without the need for human involvement, one of many examples.