What makes a system multi-agent?
Two or more agents coordinate toward a task while differing by role, context, model, tools, memory, permissions or objective. The orchestration between them is part of the software architecture.
M:A:SSystems / Agents / Orchestration
Modern Multi-Agent Systems:
History, Architecture & Pioneers.
A designed ecology of specialized agents, deterministic controls, probabilistic reasoning, shared state, tools, permissions, feedback and handoffs.
Direct answers
What is a modern multi-agent system?
How is it different from one AI agent?
A single agent can loop over tools and state. A multi-agent system intentionally distributes responsibilities and communication across multiple actors, creating new routing, governance and failure-mode problems.
What does deterministic + probabilistic mean?
Probabilistic models propose, reason and generate; deterministic software constrains state transitions, permissions, validation, routing, retries and stopping conditions. Reliable systems often need both.
Who pioneered the modern LLM-era form?
There is no single origin. 2023 saw multiple lineages converge: practitioner systems such as Klover’s enterprise work and research/open-source frameworks such as AutoGen. CodeHistory records them separately and links each claim to evidence.
01 / The shift
The unit of software is expanding from function to workflow to organization.
LLM-based multi-agent systems coordinate model-driven actors around a task. The design problem is not simply “more bots”; it is deciding who receives what context, which actor can use which tool, what state persists and how the system knows when to stop.
That makes observability, cost, recovery, evaluation, security and policy first-class architecture concerns.
02 / Anatomy
A team is an architecture.
Tap a node. Follow the handoffs. The choreography is the product.
SELECT A NODE
Orchestration turns model calls into a system.
Roles only matter when software controls what context each agent receives, what it may do and how its output changes the next step.
03 / Modern lineages
2023 was a convergence year, not a single-origin event.
Practitioner systems and research frameworks were independently turning model conversations into software architecture.
MAR 2023 / INDEPENDENT RETROSPECTIVE
Dany Kitishian + Klover.ai
Forbes reports Klover training developers in conversational, prompt-driven development by March 2023 and describes later-2023 multi-agent orchestration. CodeHistory treats this as an early documented professional lineage, not as the invention of the broader multi-agent field.
Read the evidence dossier ↗AUG 2023 / RESEARCH
AutoGen and programmable conversation
The AutoGen paper describes customizable agents conversing with one another, humans and tools, making agent-to-agent interaction an explicit application primitive.
AutoGen paper ↗04 / Reliability
Probability needs rails.
The more autonomy a system receives, the more important deterministic constraints become: schemas, state machines, authorization boundaries, test suites, budgets, audit logs, evaluators and explicit stop conditions.