-
scenarios for deploying multi-agent and agentic tools in production
- Multi-agent and agentic workflows are best suited for use cases where accuracy is the top priority and latency is not a critical or significant requirement. From performance, cost, and climate impact perspectives, running multiple LLM or SLM calls should be acceptable.
- Agentic workflows are most effective when the accuracy and quality of the output are important, but the consistency and reproducibility of the same output are not equally critical.
- The workflow should allow for natural role-play scenarios to improve accuracy and decision-making, such as two support agents discussing the applicability of a specific label in specified troubleshooting notes. While this actor-critic approach is effective, limited requirements apply for log analysis-based improvements. This is because manual or LLM-assisted analysis of role-play discussions for troubleshooting workflows can be time-consuming and tedious.
- To effectively use multi-agent and agentic tools, it is essential to have access to models with a performance equivalent to GPT-3.5 or higher, preferably LLMs or SLMs with: instruction-following ability; large context support, such as for role play and role-play summarization; the ability to consistently output JSON; and, ideally, function calling.
- Sample tasks for agentic workflows include: ground-truth generation for modeling tasks to accelerate human ground-truth compilation by reviewing agent-generated ground truth; text, video, and audio annotation for human review; offline text, video, image, or audio analysis tasks; and offline creative writing tasks such as drafting Knowledge Base articles or email messages.