AGENTPM: A SHARED BRAIN FOR TEAMS BUILDING WITH AI AGENTS

Overview

Every developer now runs AI agents locally: fast, private, and isolated. But coordination and the truth of the work are shared. Nothing connects the two. AgentPM is the layer in between: a per-workspace shared brain that holds live work-and-coordination state plus a versioned catalogue of installable capability, served identically to every human and every agent across every project.

The Problem

Coordination pain scales with people multiplied by agents, not with people alone. When several actors, human and agent, work across worktrees and machines at once, they collide, duplicate work, and drift from the standards. AgentPM exists so concurrent actors see each other before they act.

The Approach

One Django service layer sits behind multiple doors: a dashboard, the apm command-line tool, and, later, an MCP server, all resolving to the same services and tables. It is multi-tenant from the first migration (a workspace is a tenant, via django-boundary).

The design holds a few hard invariants. It is authority, not a gate: it never blocks a commit. The database owns coordination facts and files own content, so there is no document body stored in a column. And coordination facts are always live, never cached, because a stale coordination fact is worse than none.

Status

AgentPM is pre-v1, built beachhead-first: the smallest thing that proves the mechanism, dogfooded against my own workspace as tenant one. The apm CLI is already in daily use across my other projects, which bind to it for their coordination discipline. Hosting it as a multi-tenant service is a deliberately deferred later phase.

Results

• A working coordination CLI (apm) dogfooded across every project I run.
• A single service layer designed for multiple front doors: dashboard, CLI, and MCP.
• Multi-tenant from migration one, with coordination facts kept strictly live.
• Standards and installable capability modelled as one bind-not-copy versioned artefact.