Together AI Leverages AI Agents for Complex Engineering Automation

Together AI is pioneering the use of AI agents to automate complex engineering workflows, as detailed in a recent blog post. These agents are designed to handle intricate tasks such as configuring environments, launching jobs, and monitoring processes, which traditionally require substantial human oversight. By leveraging AI agents, Together AI aims to reduce manual intervention and increase efficiency in engineering tasks, particularly in the development of efficient Large Language Model (LLM) inference systems. [source]

AI Agents for Complex Workflow Automation

In the realm of coding agents, tools like Claude Code and OpenHands have demonstrated their ability to execute complex workflows. Together AI’s approach focuses on embedding these agents within an architecture that allows them to operate effectively. This involves equipping the agents with tools that facilitate their interaction with and modification of the environment, enhancing their ability to perform multi-step engineering workflows.

Key to this process is selecting tasks that are verifiable, well-defined, and supported by existing tools. Automating repetitive tasks such as infrastructure configuration and job monitoring allows human teams to focus on strategic decision-making while leaving routine operations to AI agents.

Patterns for Building Automation Agents

Together AI identifies two sets of core patterns for developing effective autonomous agents: Infrastructure Patterns and Behavioral Patterns. Infrastructure Patterns focus on building a robust agentic system environment, emphasizing the importance of good tools, comprehensive documentation, and safe execution practices. Behavioral Patterns guide the agents on how to act, including managing parallel sessions and wait times, and ensuring effective progress monitoring.

A Case Study: Speculative Decoding

Speculative decoding serves as a case study in Together AI’s approach to automation. This technique, which accelerates LLM inference by using smaller models to predict the output of larger models, exemplifies the potential of AI agents in handling complex, multi-day processes. The automation of this training pipeline has minimized human oversight and accelerated the development process.

Despite the successes, challenges remain in context management, handling novel failure modes, and optimizing resources. Together AI continues to refine its approach, aiming to expand the applications of automation to other domains such as DevOps and scientific research.

Image source: Shutterstock