Agent → Routine Catalog

The advanced workflows shipped so far were all converted from existing msp-claude-plugins subagents. This page classifies the whole agent library — every subagent in the repository — by how well it fits the Claude-managed scheduled-routine pattern: a cron, a set of MCP gateway connectors, and a prompt, with no servers to run. It is the backlog the next conversion batches are picked from.

The four archetypes

A scheduled routine is triggered by a clock, runs unattended, and surfaces a result to a human. That shape fits some agents cleanly and fights others. Every agent sorts into one of four archetypes:

• A — Scheduled reporter / auditor / scorer. Read-mostly, produces a periodic report. The cleanest routine fit: a cron is exactly the right trigger.
• B — Triage / queue manager. Makes bounded writes and uses a status transition as an idempotency guard — the way the Autotask Ticket Triage agent advances a ticket to In Progress so a rerun skips it. Fits as a routine.
• C — Investigation. Event-triggered, deep, human-in-the-loop. A cron is the wrong trigger — these run when something fires, not on a schedule. Does not fit a routine.
• D — Remediation / provisioning. Makes consequential writes and needs human input mid-task — offboarding, deployment, quoting. A poor unattended fit. Does not fit a routine.

The routine-fit verdict in each table summarises this: ✅ fits (clean — most A and B), ⚠️ fits with caveats (workable but needs scoping or a narrowed prompt), ❌ wrong fit (a routine is the wrong tool — most C and D). The connector(s) column lists the MCP gateway connectors a routine version would need; the cadence column is a suggested cron interval.

Archetype A — Scheduled reporters, auditors & scorers

Read-mostly agents that produce a periodic report. The cleanest routine fits — a cron is exactly the right trigger and the only write is delivering the report.

Archetype B — Triage & queue managers

Agents that work a queue and make bounded writes. They fit as routines when a status transition acts as an idempotency guard, so a scheduled rerun skips what it already handled.

Archetype C — Investigations

Event-triggered, deep, human-in-the-loop. These run when an alert, detection, or report fires — a cron is the wrong trigger, so they are a wrong fit for a routine.

Archetype D — Remediation & provisioning

Consequential writes that need human input mid-task — offboarding, deployment, provisioning, quoting. A poor unattended fit; a routine is the wrong tool.

Reading this as a backlog

Archetype A and B agents are the program backlog: A agents convert almost verbatim into scheduled reporting routines, and B agents convert once a status transition is identified as the idempotency guard. Future conversion batches are picked from this catalog — an archetype-A or archetype-B row, paired with a delivery adapter, is a workflow waiting to be written. Archetype C and D agents are deliberately left alone: they are real, useful agents, but a cron is the wrong trigger for them.

The catalog also records agents that look convertible but are not. datto-rmm's backup-health-monitor reads as a textbook archetype-A reporter — a scheduled backup-health rollup. It is listed under archetype D as not routine-viable today for a different reason: Datto RMM exposes no backup data. Backup and BC/DR live in the separate Datto BCDR product, which the gateway does not connect. A routine built from this agent would have nothing to read. It is a useful reminder that this catalog classifies fit — and that connector coverage, not just archetype, decides whether a workflow can actually be built.

Spotted an agent that should be the next workflow? Open an issue in the msp-claude-plugins repository.