Key Takeaways
- A real-world cost breakdown shows when to use Claude Managed Agents vs.
- running your own multi-agent infrastructure, with a clear formula to decide.
The Cost Formula
Anthropic's Managed Agents service charges $0.08 per session-hour. One session-hour is one agent running for 60 minutes of wall-clock time. Four agents running for an hour costs $0.32. This fee covers orchestration, state persistence, auto-scaling, and a dashboard, but not your separate Anthropic API usage, custom tools, or external integrations.
To decide between Managed and a DIY setup, plug your numbers into this formula:
DIY Monthly Cost = (Hardware Cost + Local Infra + Subscription Tiers)
Managed Monthly Cost = (Agent Count × Hours/Day × 30 × $0.08) + Anthropic API + Hosting
When Managed Agents Wins
Managed Agents is the cheaper and simpler option in three specific scenarios:
- Bursty, Low-Utilization Workloads: Running 4 agents for just 1 hour per day costs $9.60/month on Managed. The fixed cost of DIY hardware and subscriptions ($50/month minimum) makes Managed 5x cheaper.
- Small Scale: If you're running fewer than 5 agents, the fixed cost of your own infrastructure dominates.
- Your Time is Valuable: If managing
tmuxsessions,cronjobs, and memory limits would take 5+ hours of your time per month, the operational burden of DIY has a hidden cost that makes Managed attractive.
When DIY Claude Code Wins
Running your own agents with persistent claude code sessions can be significantly cheaper, but only under specific conditions that match the Whoff Agents case study:
- You Have Idle Hardware: A Mac mini you already own, combined with Claude Max subscriptions you're already paying for, creates a near-zero marginal cost. Managed Agents can't compete with $0.
- High, Constant Utilization: The Whoff stack runs 14 agents for about 10 hours each per day. At that scale (4,200 session-hours/month), Managed would cost $336/month just for orchestration. Their DIY hardware and infra costs only $69/month.
- You Need Deep Custom Coordination: If your workflow requires cross-process communication via Discord, Tailscale, named pipes, or custom queues that sit outside Claude's session boundary, you'll have to build that on top of Managed Agents anyway—and pay the orchestration tax twice.
How To Apply This Now
- Audit Your Current Usage: If you're experimenting with multi-agent workflows, estimate your agents' daily active runtime. Is it a few minutes or several hours?
- Check Your Hardware: Do you have a spare machine or a development Mac/PC that's on 24/7? That's sunk cost that tilts the math toward DIY.
- Start with Managed for Prototypes: For a new project with unpredictable usage, begin with Managed Agents. The low variable cost and zero infrastructure management let you test the concept. You can migrate to a DIY setup later if your usage grows predictably and exceeds the breakeven point.
- Consider the Hybrid Approach: Use Managed Agents for customer-facing, bursty interactions (like a support bot), while running scheduled, high-utilization background jobs (like data polling cron jobs) on your own persistent
claude codesessions.
The key is to run the formula with your real numbers. There's no one-size-fits-all answer, only the correct economic decision for your specific scale and constraints.
