Mastering Agentforce Architecture: Hard Limits, Subagent Ceilings, and Scalable Design

The 15-Subagent Ceiling in Salesforce Agentforce

💬 In plain words: An org can host up to 20 active agents, each agent can have up to 15 subagents (topics), and each subagent can hold up to 15 actions. These hard platform limits dictate that you cannot build a single, do-everything bot for an enterprise.

📌 Real Scenario: Meridian wanted a single agent to handle outages, billing, tariffs, meter readings, customer complaints, and new connections. Because six distinct domains required multiple specialized actions and subagents, it quickly exceeded the limit. The scalable solution is building dedicated, domain-specific agents rather than forcing everything into one bot.

Core Limits & Operational Behavior

• 20 active agents per Salesforce organization.
• 15 subagents (topics) per agent.
• 15 actions per subagent.
• 60-second timeout applied to any individual action execution.
• Standard Apex limits apply within each conversation turn (100 SOQL, 150 DML synchronous limits).
• The reasoning engine executes a ReAct (Reason-Act-Observe) loop, meaning it may invoke the same action multiple times in a single turn. Inefficient actions can rapidly exhaust governor limits.
• Agents must be deactivated prior to making modifications and reactivated once changes are saved.

Hard Limits Architecture Overview:
├── 20 Active Agents per Org
├── 15 Subagents per Agent        <-- Directly shapes solution architecture
├── 15 Actions per Subagent
├── 60s Timeout per Action
└── Apex Limits apply per Turn   (100 SOQL / 150 DML)
    └── ReAct loop execution may invoke actions repeatedly
  

🧠 Twenty, fifteen, fifteen, sixty: 20 Agents, 15 Subagents, 15 Actions, 60-Second Timeout.

🧭 360 Card — Agentforce Architecture Rules

Rule: Design for the subagent limit from day one. Build domain-scoped agents rather than a single monolithic agent.
Gain: Avoid costly refactoring and redesigns later when the system expands.
Price: More agents require independent deployment, permissioning, versioning, and monitoring.
Limits: The figures above are strictly enforced hard platform limits.
Anti-Pattern (The Monolith): Simple to conceptualize initially, but impossible to scale beyond a single business domain.
Advanced Routing: Orchestrator patterns that dynamically delegate to connected agents help manage large domains.
At Scale: In practice, routing clarity and prompt precision become critical well before hitting the 15-subagent limit.

Architectural Q&A

Q: What are the main Agentforce platform limits, and which one has the biggest impact on architecture?

🎯 Say this first: 20 active agents per org, 15 subagents per agent, 15 actions per subagent, and a 60-second execution timeout. The 15-subagent limit has the most significant architectural impact.

• The 15-subagent ceiling prevents designing a single enterprise-wide bot, requiring you to scope agents around distinct functional domains.
• Apex governor limits (100 SOQL / 150 DML) still apply within conversation turns.
• Because the underlying ReAct engine can call the same action multiple times in a single turn, sub-optimal Apex or Flow logic will trigger limit exceptions rapidly.

2-Minute Architecture Self-Check

Q1: What metadata does the Agentforce router evaluate to choose a subagent?
A1: It relies on each subagent's description—not its internal actions or step-by-step instructions.

Q2: If an agent gives the correct answer but routes through the wrong subagent, is that a pass or a fail?
A2: Fail. Subsequent turns in that interaction will inherit the wrong actions and permission context.

Q3: What is the maximum number of subagents an agent can support?
A3: 15. This hard platform limit mandates domain-specific modular architectures.