We Shipped 5 Agent Fleet Templates. One of Them Is a 6-Agent Hospital CFO.

Spinning up a multi-agent fleet from scratch takes roughly two weeks of wiring. Not the agents themselves — those you can write in an afternoon. I mean all the stuff around them: per-agent budget caps, model allowlists, PII scanning on outbound prompts, HITL rules for expensive operations, audit logging wired to a durable store, orchestration patterns to coordinate multiple specialists, and a dashboard to see what's happening in real time.

We did this work five times in three months. After the fifth time I realized we should be doing it zero more times. So we built a template gallery.

Gallery lives at curate-me.ai/templates. Five fleets, each deployable in one click, each shipping with governance preconfigured. Here's what's in them and why.

What every template ships with

Before the specifics, the common baseline. Every template comes with:

• Per-agent daily budget caps — each specialist has its own USD ceiling
• Model allowlist — agents are scoped to a specific set of models
• PII scanning — outbound prompts get regex-scanned for structured secrets
• HITL threshold — operations above a dollar amount pause for human approval
• Rate limiting — per-fleet RPM cap
• Audit trail — every LLM call and tool call logged to the dashboard

You can tweak any of these after deploying. But the defaults are the ones I'd pick if I were deploying the fleet myself — which I have, five times.

1. Customer Support Fleet

3 agents, ~$40/day, 200 RPM

The most obvious one, so I'll keep it short. A frontline support agent handles incoming Slack and email. An escalation agent picks up anything the frontline flags — and anything over a threshold has to go through a human first. An analytics agent runs on a cron and produces support metrics reports.

Why these three specifically: splitting frontline from escalation is the difference between "responds fast" and "doesn't make expensive mistakes." Without the escalation agent, the frontline has to be conservative. With it, the frontline can be fast because anything gnarly gets routed before it commits.

Governance defaults: PII scanning on, $10 HITL threshold, 3 models allowed (one cheap for frontline, one smart for escalation, one tiny for the analytics crons).

2. AI Content Team

5 agents, ~$31/day, 60 RPM

A five-agent content pipeline: researcher → writer → reviewer → analyst → publisher.

• content-researcher scans trending topics from news sources and social platforms
• content-writer produces blog posts and articles from research briefs
• content-reviewer checks drafts for quality, accuracy, and SEO
• content-analyst looks at engagement data and predicts what will perform
• content-dev handles the technical publishing via git commits

This is actually pretty close to the pipeline that writes half the posts on this blog. The lower $/day reflects that content agents spend a lot of their time reading the web, which is cheap, and only occasionally generating long-form text, which is the expensive part. RPM is deliberately low — content doesn't need to be fast.

HITL threshold is the same $10. The one that matters most is PII scanning on the writer, because prompt templates are the single biggest source of accidental data leakage in content pipelines — I've written about this before.

3. Development Team

3 agents, ~$40/day, 150 RPM

Code review, CI runner, docs writer. This fleet runs against GitHub webhooks — a PR opens, the code-review agent comments inline, the CI agent runs tests, the docs agent updates documentation if the PR changed public APIs.

The thing I'd flag about this one: HITL threshold is $15, higher than the others. Dev work has natural bursts — a big PR can legitimately cost a few dollars to review thoroughly, and you don't want the human queue firing for every review. The trade-off is that you need to trust the cost estimator. If you don't, dial it down to $5.

Governance defaults include three models: a small one for simple linting feedback, a large one for code understanding, and a medium one for docs generation.

4. Personal Productivity

3 agents, ~$30/day, 100 RPM

This is the "I want an inbox agent but I don't want it to secretly subscribe me to things" fleet. Inbox triage, calendar scheduling, web research. Nothing world-changing — what matters is the governance.

The HITL gate on this one is tuned aggressively low ($10) because personal productivity agents tend to do things that are hard to undo: responding to emails, accepting meetings, sending Slack DMs. Human approval on anything nontrivial is the sleep-at-night feature.

PII scanning is particularly important here — the inbox agent sees literally everything in your inbox, and you don't want that accidentally leaking into a web search via a poorly-written prompt template.

5. Hospital CFO Briefing

6 agents, ~$61/day, 150 RPM

This is the weird one.

A hospital CFO has to produce a monthly board-ready briefing that covers federal and state Medicaid policy, revenue cycle denial patterns, budget variances, payer mix optimization, and regulatory compliance deadlines. Traditionally this is a team of analysts spending the last week of every month assembling it.

The fleet is six specialists coordinated by an orchestrator:

• briefing-orchestrator — decides which specialists to consult, in what order, and weaves the outputs together
• medicaid-analyst — monitors federal and state Medicaid policy changes
• revenue-cycle-analyst — finds denial patterns and underpayments
• budget-analyst — tracks variances and flags anomalies
• payer-mix-analyst — optimizes payer mix and contract performance
• compliance-analyst — tracks deadlines and audit readiness

Why this template exists: one of our design partners is a hospital CFO who kept telling us "I don't want an AI coding assistant, I want the same pattern but for my actual job." So we built one. The orchestrator pattern is the thing that makes it actually work — each specialist has deep context in its narrow domain, and the orchestrator handles the meta-reasoning of what to ask whom and in what order.

The $61/day figure is for ~1 briefing per week of real depth. Running a quick briefing ad-hoc (say, a board member asks a question) costs maybe $4-8.

Governance defaults: $15 HITL threshold (briefings occasionally do deep financial lookups that cost real money), 150 RPM, two models (one for the orchestrator, one for the specialists), PII scanning on.

It's the most specialized template we ship, and I expect it to be the least-used. But it's also the most interesting proof that the orchestrator pattern generalizes beyond software engineering. Three of our design partners have asked for variants — "same pattern but for manufacturing operations," "same pattern but for compliance at a fintech," "same pattern but for university admissions." So the pattern is real. Hospital CFO just happens to be the one we built first.

What's next

We're going to ship templates for:

• Sales research fleet — prospect research, outreach drafting, CRM updates with HITL on anything sent to a real email.
• Security review fleet — SAST triage, dependency audits, CVE monitoring with HITL on anything that modifies infrastructure.
• Legal-ops fleet — contract review, regulatory monitoring, compliance deadline tracking with mandatory human sign-off on any output that leaves the fleet.

What we're not going to ship: a "general-purpose AI assistant" template. Every one of those I've seen is either useless (so safe it does nothing) or scary (so capable nobody wants it running unsupervised). The template model only works if the template has a specific job with specific policies. "Do anything" is not a specific job.

If you want a template for a specific domain, email me. The templates above came from design partner conversations — if yours fits, we'll build it.

Gallery again: curate-me.ai/templates. One-click deploy, governance preconfigured, tweak anything after.

Your harness makes the agent work. We ship the templates that make it useful on day one.