32 AI Agents on Two Droplets, One Flat Rate
How Tacavar runs 32 AI agents on two Droplets on one $50 flat-rate LLM subscription — preset teams, governor crons, wake-on-demand budgets.
Twelve agents on one team, twenty on another. Total model bill: one $50 flat subscription.
At Tacavar, we run 32 AI agents across two DigitalOcean Droplets, and the entire multi-agent system cost is a single flat rate LLM subscription per month. No per-token billing. No per-agent API keys. No watching a dashboard in fear every time an agent decides to refactor something. If you're an operator who assumed that scaling agent count means scaling API spend linearly, this post is the counter-evidence.
The Economics: Why Agent Count Doesn't Scale Your API Bill
The default mental model for agent infrastructure is per-call: every agent is a client, every client burns tokens, every token is billed. Under that model, doubling your agent count doubles your spend, and a 20-agent org chart looks like a burn-rate problem.
That model breaks the moment you swap per-token billing for a flat rate. The economics of a multi-agent system cost structure follow your architecture, not your headcount. Two levers control the bill:
- What model each agent runs. Bulk work — content drafts, code review passes, research summaries — doesn't need frontier pricing. Route it through a provider whose plan covers it.
- When agents run at all. An idle agent burning zero tokens costs zero dollars, even at flat rate. The trick is making idleness the default state, not the exception.
At Tacavar, both teams — Bailian (12 agents) and Paperclip (20 agents) — run on Alibaba Coder Pro, a $50/month flat plan. Most bulk agents execute through qwen3-coder-plus, which the plan covers. The premium-model agents are few, and their runs are short and budgeted. This is how you run AI agents cheap: not by throttling your agents into uselessness, but by making sure the expensive ones only speak when it matters.
Bailian: A 12-Agent Team Driven by One Preset JSON
Bailian is the Tacavar growth-and-engineering swarm. Twelve specialized agents: a CEO, a Research Lead, a Dev Lead, Dev-1 and Dev-2, Frontend, an SEO Manager, a Marketing Lead, a Content Writer, a Grok Researcher, a Data Analyst, and a GPT Analyst. They're organized into two swarms — Engineering and Growth — so the org chart actually mirrors how a small company divides labor.
The entire team is defined by a single preset JSON file. One file assigns each agent its model. Change the file, restart the service, and the team runs on a different mix. There's no config sprawl, no per-agent YAML buried in twelve directories. The preset file is the team.
That design choice pays for itself constantly. When we want to test whether a Claude-led engineering swarm outperforms a qwen-led one on actual Tacavar work, we don't rewrite infrastructure. We swap a preset and issue one systemctl restart. More on that below.
Paperclip: 20 Agents and a Heartbeat Governor Cron
Paperclip is the larger operation: 20 agents including a CEO, code-reviewer, content-writer, three devs plus a dev lead, devops, qa-tester, app-dev, game-dev, frontend, seo-manager, grok-researcher, marketing, research-lead, pentest-lead, data-analyst, ui-designer, a bailian-executor, and a gpt-analyst. Twenty processes that would look like an API-billing horror story under per-token pricing.
All 20 run through a single shim — remote-qwen-shim.sh — against Alibaba Coder Pro's flat $50/month. Every agent speaks to the same subscription. No per-agent keys, no separate quota tracking.
The load-bearing component is a heartbeat governor cron that fires every 60 seconds. It's the only thing that decides when agents wake and when they sleep:
- New issue assigned? The governor wakes the relevant agent.
- Nothing to do? Agents sleep.
- The CEO agent is the exception — it stays always-on with a 5-minute heartbeat, because the coordinator's job is to notice things the others can't.
This is the agent heartbeat pattern in its simplest, most operational form: one cron, 60-second cadence, wake-on-assignment, sleep-on-idle. It replaced what would otherwise be 20 agents polling constantly and burning plan quota on nothing. If you're doing agent orchestration as a solo founder, the governor cron is the single highest-leverage piece of infrastructure you can build.
Wake-on-Demand: Sleeping Agents and Per-Agent Turn Budgets
Sleeping agents solve the when problem. Turn budgets solve the how much problem.
Every Paperclip agent has a per-agent max-turn budget, enforced at the runtime level. The CEO gets 50 turns — it coordinates, delegates, and needs room to think across the whole operation. The dev lead gets 35. The content writer gets 30. Everyone else runs at 20.
These aren't arbitrary numbers. They encode the operational reality that some roles need long autonomous chains and most don't. A qa-tester that needs 50 turns to verify a change is either under-specified or out of control. The budget caps make worst-case spend predictable: even if every agent fired simultaneously and maxed out, the plan consumption is bounded and known.
The budgets also make failure cheap. An agent that starts looping hits its cap, the run ends, and the governor simply won't wake it again until there's new work. A runaway loop under per-token billing is an incident; here it's a capped line item.
Combined with wake-on-demand, the math changes completely. Agents that sleep when there's nothing to do — which is most of the time — consume almost nothing. The 20-agent Paperclip team spends the majority of its life dormant, waking in bursts when issues land. That's why multi-agent system cost doesn't scale with agent count here: it scales with work volume, and work volume is what a $50 plan can absorb.
Preset Switching: Rebalancing the Whole Team in One Restart
Model assignment at Tacavar is data, not code. Bailian's preset JSON supports several mixes — qwen, claude, gpt, and a best_mix composite — and switching between them is a file edit plus one systemctl restart.
Why this matters operationally:
- A/B testing model mixes. Run a week on qwen-led, a week on best_mix, compare shipped output. The experiment costs a restart, not a refactor.
- Failover without panic. When a provider has an outage or a rate problem, route the bulk agents to a different covered model and keep operating.
- Cost re-balancing in seconds. If a premium-model agent is burning quota faster than expected, demote it in the preset and promote a plan-covered model. The whole team rebalances at once.
For a solo operator, this is the difference between infrastructure you can tune and infrastructure you're stuck with. Twelve agents' model assignments live in one JSON file that fits on one screen. That's the entire model-routing layer.
The Flat-Rate Playbook for Solo Operator Agent Teams
Everything above reduces to a repeatable playbook. If you want to run AI agents cheap with a team this size, the sequence is:
- Pick one flat rate LLM subscription as your backbone. One provider, one plan, one bill. Bulk work goes here — full stop.
- Shim every agent through it. Paperclip's
remote-qwen-shim.shmeans 20 agents share one endpoint. One integration, not twenty. - Build the governor before you scale headcount. A heartbeat governor cron on a 60-second cadence, wake-on-assignment, sleep-on-idle, with one always-on coordinator. The agent heartbeat pattern is what makes 20 agents cost the same as 2.
- Budget turns per role, not per agent. 50 for the CEO, 35 for the dev lead, 20 for specialists. Encode role seniority into the runtime.
- Keep model assignment in one preset file. If changing your team's model mix requires touching more than one file, you'll never experiment — and experimentation is where the savings are.
Two Droplets, one flat subscription, 32 agents. The Tacavar setup proves that agent orchestration for a solo founder isn't a spending problem — it's an architecture problem. Agent count is free. Idle time is free. Tokens are what cost, and tokens are what the flat plan and the heartbeat governor jointly minimize.
Meet the Bailian and Paperclip agent teams running Tacavar at tacavar.com/agents.