Multi-agent pipeline (retired)
Last Updated: 2026-04-23 Status: RETIRED. This document’s previous content described a pre-Phase-E architecture (6 named agents with a self-critiquing loop, Financial / Market / Compliance specialized agents, a BaseAgent hierarchy) that no longer matches the shipped code.Where to look now
The Phase E refactor replaced the agent-centric model with a stage plugin chain, which Lane C (2026-05-10..16) further evolved into a LangGraph-based template (canonical_blog, currently 39 nodes — a
DB-stored graph_def as of atom-cutover #355; the 5-node dev_diary
template lives alongside it). The plugin roadmap
locked in Phase J added an LLMProvider family that pluralizes
inference backends. All current and future pipeline work is documented
in:
- Template Runner — the authoritative
description of the LangGraph orchestrator that drives the 42-node
canonical_bloggraph_def (and the 5-nodedev_diarytemplate) that runs every content task today. Covers node behavior and halt semantics. - Plugin architecture — the evolution plan from god-files to plugin Protocols (Tap, Probe, Job, Stage, Pack, LLMProvider). Umbrella issue GH-64.
- Services reference — catalog of
every service in
src/cofounder_agent/services/with the four blog-focused agents (content generator, image, publisher, quality) called out under Core and Pipeline orchestration. - Database schema — every table + migration.