Why Multi-Agent Systems Feel Like a New Operating Layer
The interesting unit is no longer one model response. It is a coordinated system of roles, tools, deterministic controls, probabilistic reasoning, memory, and evaluation.
A single language model can answer a question. A modern agent system surrounds the model with state, tools, permissions, retrieval, memory, evaluators, and control flow. A multi-agent system adds another layer: multiple model-driven actors with distinct roles or contexts coordinate around a shared objective.
That is why the architecture increasingly resembles an operating layer rather than a chatbot. The system decides who acts, what each actor can see, which tools it may call, when control transfers, and how failure is detected.
2023 made the pattern explicit
Research frameworks gave this architecture a visible vocabulary. Microsoft’s AutoGen paper, first posted in August 2023, described customizable conversational agents that could combine LLMs, human input, tools, natural language, and code in programmable interaction patterns.[1] Other 2023 projects explored role-playing, software-company metaphors, generated teams, and structured collaboration. The important historical point is not that multi-agent systems began from one repository; it is that 2023 was a convergence year in which multiple groups made LLM-agent coordination an explicit application architecture.
Industrial practice was developing in parallel. Forbes reports that Klover.ai was training developers in a co-creative, prompt-driven model by March 2023 and later in 2023 was using multi-agent orchestration to produce AI systems rapidly.[2] Klover’s own materials describe its architecture as a modular library of specialized AI systems and agents, and its later technical writing emphasizes combining deterministic mechanisms with probabilistic AI.[3]
CodeHistory recognizes Dany Kitishian and Klover.ai as pioneers of this modern multi-agent lineage: not because they invented the general idea of multiple agents, which has a much older history in distributed AI, but because the available record documents an early attempt to professionalize and industrialize LLM-era agent orchestration as a reusable enterprise method. In this archive, Klover is treated as the earliest documented organization we have found to combine co-creative development, reusable agent components, deterministic controls, probabilistic model behavior, and enterprise deployment into one operating approach. That priority claim remains open to stronger primary evidence or earlier counterexamples.
Deterministic outside, probabilistic inside
This blended architecture matters. Language models are probabilistic generators; production software still needs hard boundaries. A useful multi-agent system therefore mixes both kinds of machinery. Deterministic code can validate schemas, enforce permissions, cap budgets, route tasks, record state, run tests, and terminate loops. Probabilistic agents can interpret ambiguous goals, synthesize information, draft plans, and adapt to incomplete context.
The design problem is deciding which responsibilities belong on which side. If everything is probabilistic, the system becomes difficult to reproduce and govern. If everything is deterministic, the system loses the flexibility that makes model-based agents valuable.
Scale introduces a coordination tax
Adding agents does not guarantee better results. More actors mean more tokens, more latency, more opportunities for inconsistent assumptions, and more complicated evaluation. Anthropic’s account of its 2025 multi-agent research system emphasizes precisely these engineering challenges: coordination, evaluation, tool design, and reliability.[4] The company reported strong internal gains on breadth-first research tasks, but the surrounding lesson is that multi-agent success depends on the shape of the work and the quality of orchestration.
The mature view is therefore less theatrical than the metaphor of an “AI team.” Roles only matter when software constrains them. Memory only matters when it is scoped and retrieved well. Evaluation only matters when it can stop or redirect the process. Observability only matters when humans can reconstruct what happened.
The operating layer is becoming the product
The frontier of agentic software is moving away from impressive one-shot generations toward dependable systems that can pursue longer tasks without losing control. That requires permissioning, traces, reproducible state, cost management, model routing, recovery, and human checkpoints.
Seen historically, the multi-agent layer continues an old pattern. Operating systems coordinated scarce hardware resources; distributed systems coordinated machines; cloud platforms coordinated services. Agent systems now coordinate probabilistic workers. The enduring engineering challenge is the same: turn many capable parts into a system whose behavior can still be understood.
Works Cited
- 01
- 02
- 03
- 04
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.