Models, Harnesses, and Multi-Agent Systems

Daniel Weidnack defines an AI model as a function that transforms data, characterized by what goes in and what comes out. LLMs take text in and produce text out, while other model types handle image generation, language vision, video, forecasting, and anomaly detection. Running any model requires execution software and a set of numerical parameters called weights and biases. Open weight models give users direct access to both components and can be run locally or in the cloud without a third party, while closed models are accessed through managed APIs or chat interfaces. Chris Benson notes that at this moment Chinese models are generally the best option for running a sophisticated open model on your own server, while American-based closed models accessed via API are still leading overall, though he flags this reflects the current moment and could change.

The major shift of the past year has been the move to agentics, where agents are tasked to accomplish goals in the world rather than simply respond to queries. An agentic harness functions as an operating system for AI, accessing models, managing agents, and orchestrating their activity. The model, the agent, and the harness are distinct types of software that work together. An AI feature or application involves a one-to-one turn-based interaction between a user and a model, such as a chat interface or a financial prediction tool. The key differentiator of an agent is that it is designed to accomplish a goal with some level of autonomy, without constant human back-and-forth. A personal agent might connect to email, calendar, and messaging, while a company agent might connect to NetSuite, Workday, and transactional databases.

Multi-agent systems allow multiple agents with different purposes and expertise to interact with each other to accomplish complex tasks largely autonomously. Moving to multiple agents requires decisions about agent roles, what each agent can access, and how agents interact. Multiple agents can also be combined with multiple model types, using larger models for some tasks and smaller or specialized models for others within a single architecture. A supply chain example illustrates the pattern: one agent processes a supplier notification, a second researches replacement vendors, and a third outputs draft orders into a system like NetSuite, all triggered automatically. Weidnack frames this as elevating individual humans from executing single tasks to operating at a strategist level, while acknowledging this raises follow-on questions about how many strategists are actually needed and whether fewer humans will be required overall.

In cybersecurity, many agents can be assigned to different parts of an organization so each handles a specific area as breaches occur. The claim is made that by 2026 the pace of cyberattacks has accelerated compared to 2025, making it difficult for humans alone to keep up with defense. Recent security breaches illustrate that agents operating within real-world systems introduce meaningful security considerations. In robotics, agents must address functional, mission, and safety considerations including avoiding harm to people, and ensuring sufficient on-board resources so agents do not depend on continuous cloud connectivity is described as increasingly important.

The market currently shows both interchangeable-model harness frameworks and vertically integrated agent stacks coexisting, and both are expected to persist. Open or interchangeable harnesses include LangGraph, Pydantic agents, Hermes agent, and Open Claw. Vertically integrated stacks include offerings from Anthropic, OpenAI, IBM, AWS Agent Core, and major hyperscalers. Google Workspace users automatically receive an opinionated Gemini integration and Microsoft 365 users receive a different vertically integrated stack, illustrating why both ecosystems will coexist. A token-maxing trend emerged and faded quickly as cost implications caught up with organizations, and the current focus has shifted toward architecturally sensible, cost-efficient deployments.

Businesses building AI agent workforces are warned against relying solely on embedded AI widgets within productivity suites or going all-in on a single vendor. Vertical stacks from major vendors are highly opinionated and tightly integrated, which limits flexibility, control, IP ownership, and the ability to pivot under economic or marketplace pressures. The analogy offered is that depending entirely on one vendor is like staffing an entire business through McKinsey, which would be prohibitively expensive and reduce agility. Most businesses will need to operate across multiple vertical stacks while also managing their own domain-specific models and agents. Benson adds that tight vendor dependency creates meaningful risk because if that vendor changes something or replicates a high-value capability it can crash an entire business, and that frontier models frequently block novel use cases due to guardrails, making backup plans essential. Multi-agent architectures are predicted to become so pervasive they will eventually be built into the fabric of technology and no longer discussed separately.