CIYL Engine — a multi-agent content production workflow

Context

Producing structured editorial content — analyses, briefs, reference texts — follows a repetitive pipeline: gather the material, write, verify, format. Done by hand, the pipeline is slow and uneven. Done by a single agent, it concentrates too many responsibilities in one place.

Problem

A generalist agent provides neither separation of responsibilities nor an explicit quality gate: the same context produces, verifies and corrects. Errors propagate with no checkpoint.

Constraints

• Consistent editorial quality from one issue to the next.
• Traceability: knowing which step produced what.
• Controlled cost (tokens, calls, processing time).
• The same mechanics adapted to several editorial domains.

My role

Pipeline architecture, design of the steps and their input / output contracts, the verification loop and the human checkpoint.

Approach

Break the workflow into specialised steps — research, writing, fact-checking, quality control, formatting — instead of one generalist agent. Each step receives a bounded context and produces a typed deliverable. Verification never reuses the production context: it starts from the deliverable.

Architecture

• An orchestrator that sequences the steps and carries the global state of each dossier.
• Specialised agents per functional role (research, writing, verification, formatting), with explicit input / output contracts.
• Persistent context between steps, limited to what is strictly necessary.
• A verification loop independent from production.
• A human checkpoint before publication.

Key decisions

• Specialisation over generalism: one agent per role, a bounded scope, a typed deliverable.
• Production and verification separated: verification starts from the deliverable, never from the writing context.
• Human approval at critical points: the pipeline proposes, the human publishes.
• Domain as configuration: changing the editorial topic does not change the mechanics.

Challenges

Cross-step consistency: every specialisation risks losing context at the boundary. Keeping instructions from drifting and keeping the per-dossier cost under control requires instrumenting the steps, not tuning by feel.

Result

A working content production pipeline, deployed across several editorial domains, with a steady cadence and human approval maintained.

What this project demonstrates

Designing useful multi-agent workflows: specialise, separate, trace, and keep humans where they matter — rather than piling up autonomy.