2760 - 为企业用例高效构建开源代理.pdf

1、IBM Data&AI|2025 IBM CorporationEffectively Building Open Source Agents for Enterprise Use Cases 2760Mansura Habiba,IBMAI capabilities are growing rapidly AI that predictsAI that predicts Machine learning AI that createsAI that creates Generative AI AI that chatsAI that chats AI assistants AI that w

2、orksAI that works AI agentsAI agents will drive the next waveof knowledge worker productivity25%decrease in time to complete tasks when using AI tools112%more tasks completed in total with the help of AI tools140%higher quality in work completed with AI tools1By 2028,1/3 of interactions with generat

3、ive AI will use autonomous agents for task completion.22.Gartner 2024 predictionBuilding Agents is hard Many different solutions What models to pick What tools to use What pattern to useMany different solutions Open-source agent frameworks Visual agent builders Enterprise multi-agent orchestration t

4、oolsMany different solutions Open-source agent frameworks Visual agent builders Enterprise multi-agent orchestration toolsThings to check for model SelectionTool Use Capability for structured outputs&function callingModels like Granite,Llama,GPT-OSS etcReasoningCapability to support multi-step plann

5、ingContextRAG pipelines for enterprise data7Latency&ScaleOptimized GPU pools,lighter 7B models for speedGroundednessRAG+Granite fine-tuned for factual enterprise useGovernancebias,risk&audit controlsDeploymentCloud,on-prem(OpenShift),or hybrid flexibilityCostOSS avoids licensing;quantization lowers

6、compute spendExtensibilityFine-tune with LoRA;integrates with LangChain/LangFlowWhat models to pick Open-source models on watsonx.aiWhat tools to use Each framework comes with own way of specifying tools How will tools orchestrate together?How does the accessibility of tools,including security,authN