A Three-Layer Agent Ecosystem for Project Managers
A personal system of orchestrator agents, specialised skills, and a local-first knowledge engine that turns scattered PM work into an orchestrated, memory-aware practice.
The problem
A project manager's day is meetings, decisions, documents, communications, and context-switching. Most of that work is gathering information, synthesising it, and acting on it. The tools most organisations use each handle one slice well. None of them hold the thread across all of it.
I started with the most repetitive workflow, meeting minutes, and built one narrow agent to handle it. It worked. So did the next one. The ecosystem grew from there, one verified workflow at a time, entirely outside work hours.
Architecture
- Meeting processor
- PM assistant
- Dev assistant
- Morning sentinel
- Communications
- Process knowledge
- Change requests
- Business cases
- Meeting prep
- Project context
- Planning
- Coaching
- Slides
- …and more
- PostgreSQL + pgvector RAG
- Hybrid semantic + full-text search
- 4-tier memory
- Bitemporal memories
- Markdown knowledge vault
- Background LLM worker
Agents orchestrate. They run the long, multi-step workflows, pause at review gates, and never write anything back without confirmation.
Skills transform. Each does exactly one thing, has no state of its own, and returns structured output. The communications skill does not manage tasks. The meeting processor does not draft emails. Narrow beats smart.
Infrastructure remembers. A local-first retrieval engine over PostgreSQL and pgvector, hybrid semantic and full-text search, and a four-tier memory model: working, short-term, long-term, and core. Every decision, correction, and meeting is searchable and time-aware.
Design principles
Spec first. Every component starts as a written design. Implementation plans are reviewed before code.
Every write needs a gate. Nothing reaches a project record, a document store, or long-term memory without a dry run and an explicit confirmation. Not paranoia. PM discipline applied to AI.
Shared protocols, not copied logic. Thirty-five protocols cover session lifecycle, retrieval patterns, self-critique, writing style, and a set of composable thinking moves such as pre-mortem, steel-man, and second-order analysis. Skills reference them rather than re-implementing them.
The interesting part is not the AI. It is that every design decision came from PM experience. The agents are narrow because general-purpose tools fail in practice. They show their reasoning because trust requires transparency. They ask before they write because I have seen what an unreviewed change does to a project.
Self-evolving. The system has its own development assistant that reviews plans, runs retrospectives, and checks architectural consistency. The ecosystem maintains itself with the same discipline it applies to project work.
What it changed
Meeting preparation that used to mean an hour of searching now starts from a brief that already knows the history, the open risks, and the pending decisions. Communications adapt to the recipient. Process questions are answered from the organisation's actual procedures rather than from memory. The work did not disappear. The gathering did.
Boundaries
This is a personal project, built and run outside work hours. Where it touches an employer's processes, it is grounded in general procedure knowledge only. No internal systems, data, or tools are described here.