开源与人工智能的未来：智能体如何颠覆我们的系统、先例以及软件中的人类角色

How Agents are Disrupting OurSystems, Our Precedent, and theHuman Role in Software Hilary Carter, The Linux FoundationAnna Hermansen, The Linux Foundation April 2026 Open Source and the Future of AI To build trust betweenindividuals and the agentsacting on their behalf,users need the ability toset fine-grained The programmer’s role isevolving intoan architectwho designs and definesproblemswhile delegating The success of open source AIinfrastructure such as Rayand vLLM demonstrates three While developers are movingquickly to grant agents APIkeys and access,essentialsafeguards are almost Without clear rules onaccountability for agentbehavior or a unified processfor asserting identity, Reasoning traces in openmodels are integral tosecure adoption, allowingusers to inspect decision paths Open source acts againstvendor lock-in and singlepoints of failure, ensuring theflexibility to swap agentsandcustomize workflows Before an agent canautomate human workflows,organizations must provide it Human accountabilitymust remain the finalstamp of quality for Open source projects areactively solving AI’s mostpressing challenges bysupportingspecialized AI Future scaffolding for agentsincludeslicenses, openevaluations, andcommunity-driven projects The AI ecosystem mustestablish acomprehensiveframework of legal Contents Executive summary Third, in regulated industries like banking and healthcare,the challenge of adoption is compounded by the need todocument complex human processes that agents are expectedto automate in a workflow. Because a computer cannot be heldlegally accountable, humans in these sectors must be diligentabout teaching their agents and ultimately retaining the finalstamp of quality to ensure compliance and risk management. The AI Executive Forum, held on February 26, 2026, in SanFrancisco, brought together leaders from industry and opensource projects to discuss how autonomous agents aredisrupting technical systems and the human role in software.Hosted by the Linux Foundation, the event took place at acritical point in the adoption and use of agentic systems thatcan independently execute complex workflows. The forum The second half of the forum consisted of 4 roundtableswhere participants discussed major challenges that AI agentspose to the technical community today. First, as these agentsbecome active economic participants, a trust mandate hasemerged around their identity and accountability. Participantsemphasized that delegating identity to agents requires fine-grained control and authorization limits to build confidencein autonomous actions. Without standardized systems for The infrastructure supporting agentic transformation is builtupon several critical open source projects that act as a controlplane for enterprise AI. These include the Model ContextProtocol (MCP) for connecting models to data, PyTorch forresearch and isolated execution environments, Kubernetes fororchestrating AI-first hardware, Ray for distributed computecoordination, and Goose for local-first, private agenticexperimentation. To capitalize on this innovation and growth Introduction As we enter 2026, the artificial intelligence (AI) landscape isshifting and evolving at increasingly rapid speeds, while at thesame time industry leaders and technical project ecosystemslay the groundwork for what is quickly becoming mission-critical technology. At the heart of this transformation are twodefining forces: open source AI and AI agents. While open the forum. What follows is an exploration of current and futurestates of open source AI and programmers, major challengesthis ecosystem is currently facing, and the foundationalopen source projects that are sustaining the ecosystem andworking to address some of these challenges. Drawing from This transition brings significant challenges as well asopportunities. The ecosystem is navigating growth intohigh-stakes business environments. Organizations face adelicate balance: the need for rapid innovation that createsefficiency on the one hand, and the need for trust, security,and compliance on the other hand. To support theseneeds and provide a neutral home for development that is On February 26, 2026, the LF brought leaders from theseprojects, as well as from major enterprises, startups, academia,and foundations, together for the AI Executive Forum, an The past, present, and future states The forum began with a plenary session to introduce majorthemes and perspectives of the day. AI thought leaders spokeon lessons learned, current challenges, what changes are same hardware and software infrastructure. It became thefoundation for the first OpenAI models, and its success was The growing number of LLM-powered services has led toincreased model serving, but the size and general nature ofthese LLMs make them very expensive to serve. The solutionbecame batching requests, but this approach soon strainedthe memory of the model. Ultimately, vLLM introducedPagedAttention to eliminat