← back
Cohere: Building enterprise LLM agents that work (Shaan Desai)
Takeaway
Enterprise agents that actually work come from disciplined tool-spec engineering, conservative multi-agent design, and step-level evaluation—not framework choice.
Summary
- Cohere's Shaan Desai evaluates agent frameworks on three axes: observability, setup cost, support—recommends native/LangGraph for enterprise, CrewAI/AutoGen for POCs.
- Start simple: one LLM + few tools beats complex multi-agent setups when tool specs are tight (clear descriptions, flat input types instead of nested dicts, short instructions).
- Multi-agent guidance: router needs sharp routing instructions covering edge cases; sub-agents constrained to independent tasks with small toolsets.
- Caching long chat histories (>20 turns hallucinate), human-in-the-loop gates before/after risky tool calls, and step-level eval (not just final answer) are critical.
agentscohereenterprise
Original description
Building scalable, safe and seamless LLM agents for enterprise is a difficult task. Not only must developers choose the right frameworks, tools, and models from a panacea of options but they also need to carefully orchestrate them to build fault tolerant agents that meet strict enterprise requirements. In this talk we walk you through the critical decision making process in setting up enterprise agents. Specifically, we discuss the frameworks we love, how we define robust tools and ensure safety via human-in-the-loop, highlight our evaluation criteria and importantly, discuss how we improve model capabilities.