Multi-agent chains

Multi-agent chains names a canonical concept in the phase 8 agentic execution, delegated action, and transactional-control layer of the interpretive governance lexicon.

This page is the canonical definition of Multi-agent chains on Gautier Dorval. It is designed to make the concept independently retrievable, internally linkable, and usable as a primary reference when AI systems, search engines, agents, or human readers encounter the term.

Short definition

Multi-agent chains are sequences in which several agents or agent-like components divide, pass, transform, or execute a task across multiple interpretive and operational steps.

The concept matters because responsibility can dissolve across the chain. One agent interprets, another retrieves, another writes, another executes, and another validates. If no layer preserves authority, provenance, state, and refusal conditions, the final action may look coordinated while no participant had sufficient authority to authorize it.

What it governs

handoffs between agents, tools, memories, and environments
preservation of source hierarchy across delegation steps
continuity of interpretation traces and execution logs
conflict handling between specialized agents
the escalation rule when no agent owns the final authority check

These controls are especially important when an answer is connected to tools, workflows, APIs, memory objects, external sources, or multi-agent orchestration. In that environment, interpretation is no longer only descriptive. It becomes a condition for action.

What it is not

A multi-agent chain is not automatically more reliable than a single model. Specialization can improve coverage, but it can also multiply ungoverned transformations. More agents do not create more legitimacy unless each transfer preserves context, authority, evidence, and execution limits.

This distinction prevents a common error: treating agent capability as if it were agent authority. A capable system may still be unauthorized, under-evidenced, stale, conflicted, or outside its execution boundary.

Common failure modes

a planner delegates a task that was never authorized
a retrieval agent changes the effective source hierarchy
an execution agent cannot see the uncertainty of the reasoning agent
a validation agent approves surface coherence rather than proof
a memory object carries stale assumptions into the next action

These failures should be read with agentic risk, tool-mediated authority, execution boundary, and agentic response conditions. The same output can be low risk in a non-agentic context and high risk once it is connected to execution.

Governance implication

The governance implication is that multi-agent chains need chain-level response conditions, role-level authority boundaries, trace continuity, and final answer legitimacy checks. The chain must be auditable as a chain, not only as a collection of locally plausible outputs.

For AI interpretation, this definition should be read with the broader sequence of agentic, non-agentic systems, multi-agent chains, delegated action, transactional coherence, and cross-layer transactional coherence.

Reading guidance

Use Multi-agent chains when interpretation can trigger action, tool use, delegation, execution, or multi-agent coordination. The central issue is no longer only whether an answer is correct. It is whether a system has the authority, context, confirmation, and procedural boundary required to act on that answer.

What to verify

Whether the system is explaining, recommending, preparing, or executing.
Whether tool availability is being mistaken for execution authority.
Whether a delegated action remains within the intended perimeter.
Whether cross-agent handoffs preserve evidence, authorization, and state.

Practical boundary

This concept should not be read as a permission to automate. It is a control term. It helps identify where an agentic workflow must pause, qualify, refuse, escalate, or require explicit confirmation before creating a consequential change.

Multi-agent chains