Source: ai-research (web research, 2026-04-11) Type: Comparison / Decision Framework
Claude now has three distinct agent tiers: Managed Agents (hosted infrastructure), Agent Teams (collaborative multi-instance), and Subagents (isolated parallel workers). They solve different problems and operate at different levels of complexity. This article maps out when to reach for each one.
The three tiers at a glance
| Managed Agents | Agent Teams | Subagents | |
|---|---|---|---|
| Where it runs | Anthropic’s infrastructure | Local (Claude Code / Cowork) | Local (Claude Code / Cowork) |
| Setup effort | Minimal — define in YAML/NL | Enable experimental flag | None — built in |
| Agent communication | Orchestrated by platform | Peer-to-peer between teammates | No inter-agent communication |
| Persistence | Long-running, checkpointed | Session-scoped | Task-scoped |
| OAuth/credentials | Built-in | Manual | Manual |
| Cost model | $0.08/hr + tokens | Standard Claude usage | Standard Claude usage |
| Status | Beta (April 2026) | Experimental | Stable |
Decision framework
Use Managed Agents when:
- The agent needs to run continuously or on a schedule (monitoring, polling, recurring tasks)
- You need OAuth integrations with third-party services without building the plumbing
- You want to avoid infrastructure work entirely — no sandboxing, checkpointing, or credential management
- Long-horizon tasks that may take hours or days to complete
- Example: “Monitor this Slack channel and file a Linear ticket whenever someone reports a bug”
Use Agent Teams when:
- Multiple agents need to share discoveries in real-time during a task
- Tasks require agents to challenge each other’s work (review while implementing)
- Cross-cutting changes where one agent’s decisions affect another’s work
- You want to interact with individual teammates directly, not just through a coordinator
- Example: “Refactor the API — one teammate handles the backend endpoints, another handles the frontend integration, and they coordinate on the contract”
Use Subagents when:
- Tasks are independent and parallelizable — no need for agents to talk to each other
- You want to keep the main context window clean by offloading research or analysis
- Different tasks need different permission levels (read-only auditor vs. code writer)
- You need reusable agent definitions stored in
.claude/agents/ - Example: “Run a security audit on the auth module while simultaneously generating test coverage for the payments module”
Combining tiers
These tiers are not mutually exclusive:
- A Managed Agent could be the always-on orchestrator that triggers Subagent tasks when events occur
- An Agent Team working on a large refactor might spawn Subagents for independent research subtasks
- Cowork (marketing) uses the Agent Teams architecture for parallel creative work + Subagents for research
Cost comparison
- Subagents: Free beyond standard token usage. Best for cost-sensitive parallel work.
- Agent Teams: Free beyond standard token usage. Slightly higher token consumption due to inter-agent communication.
- Managed Agents: 58/month baseline for 24/7. Worth it when infrastructure savings exceed the cost.
Key Takeaways
- Subagents = independent parallel workers. Cheapest, simplest, most stable. Default choice for parallelism.
- Agent Teams = collaborative workers that talk to each other. Use when coordination matters more than independence.
- Managed Agents = hosted infrastructure with built-in OAuth, checkpointing, and persistence. Use for long-running or always-on agents.
- Start with subagents. Upgrade to agent teams when you need peer communication. Move to managed agents when you need hosted infrastructure.
- The three tiers can be combined — a managed agent can orchestrate subagents, and agent teams can spawn subagents for subtasks.
Try It
- Start simple: Use a subagent for your next parallel research task in Claude Code — “have a subagent investigate X while I work on Y”
- Try agent teams: Enable the experimental flag, then give Claude Code a task that requires two specialists to coordinate (e.g., frontend + backend changes to the same feature)
- Evaluate managed agents: Identify a task you currently run on a cron job or manually trigger — estimate if $58/month + tokens is cheaper than your current maintenance cost
- Combine: Design a workflow where a managed agent monitors an external service and spawns local subagents when action is needed
Related
- Claude Managed Agents
- Claude Code Agent Teams
- Claude Code Subagents
- The Advisor Strategy (advisor_20260301) — fourth pattern: upward advisor consultation (inverts sub-agent decomposition)
- Claude Cowork for Marketing
- AI Marketing Automation Use Cases
- The Complete Guide to Building Skills for Claude