glossary//multi-agent-pipeline

frameworks

multi-agent pipeline

A multi-agent pipeline sequences distinct agent roles that pass artifacts forward through gated handoffs.

also: agentic pipeline, multi-agent workflow·Sep 2026

The word doing the work here is structure. A pipeline is the shape, not any of the roles inside it. The orchestrator is the single role that runs a pipeline, so the two get used interchangeably and should not: one is the route, the other is the driver. A critic or a gatekeeper is a station on the route.

It is also not one agent looping on a task until it looks done. The defining property is that the steps are different roles, in separate contexts, passing artifacts forward. That separation is what makes a later reviewer's objection worth anything, because it never watched the earlier reasoning happen.

What the chain looks like

Gate the flow, not the judgment sketches a build pipeline in one line: an analyst frames the problem, an architect turns it into a plan, a critic tears holes in the plan, a prompt-engineer scopes the work, and a fleet of coders does it. Between every stage there is a handoff, and at every handoff something decides whether to proceed, send it back, or stop and escalate.

Shift-left quality in an agentic dev pipeline describes the same chain from the reviewer's side: an Analyst produces requirements, an Architect turns them into a plan, a Scribe turns the plan into prompts, coders execute the prompts, a Critic reviews the outputs. The interesting question in that essay is not what the Critic catches but when, which is a property of the pipeline's shape rather than any agent in it.

A pipeline that is not writing code

Nothing about the pattern is specific to software builds. The pipeline that wrote this documents an authoring chain, roughly fifteen steps once you count the gates and actions between them: frame, outline, draft, revise, ground the figures, a cross-model editing pass, reconcile, two humanizer passes, a cold proofreader, publish.

What makes that runnable without a human at each handoff has nothing to do with model quality. It comes down to immutable stages you can diff at any point, a cold reader that sees only the final file, a check that traces every number back to where it came from, and an edit from a different model entirely. The structure around the model does the trustworthy part.