#Enterprise Architecture8 utenti parlano di questo argomento

Hey Jiyaulla,

Here’s how I would approach the solution design:

1. Agent Boundaries & Routing

• Split by persona: Don't build one monster agent. Keep Employee Desk (IT/HR Q&A via Knowledge), Coworker (sales/service reps on record pages), and OOTB Service Agent in clean, separate buckets.

• Topics & Hand-offs: Limit each agent to 4–6 tightly scoped Topics. Write explicit "When to use" instructions so the reasoning engine doesn’t cross wires. Always map human escalation via Omni-Channel, carrying the chat summary over.

2. Actions & Grounding (The Engine)

• Actions: Use Autolaunched Flows for standard CRM updates and Invocable Apex for external API lookups. Keep JSON outputs short; feeding raw multi-page payloads burns tokens and slows reasoning down.

• Flex Prompt Templates: Stick to a 4-part prompt: Role, Grounded Context, Guardrails (what NOT to do), and exact Output Format.

• Knowledge & Data Cloud: If grounding with PDFs or policies, ensure your Data Cloud search indexes chunk properly. Watch Field-Level Security—if the running agent user can’t see the field, the LLM hallucinates.

3. Trust, Guardrails & Security

• Einstein Trust Layer: Decide upfront if PII needs data masking and where audit logs get retained.

• Hard Refusals: Put explicit negative prompts inside each topic (e.g., "Never disclose salary bands; route user to Workday").

4. Licenses & Consumption Reality Check

• Cost control: Agentforce bills per conversation, and custom prompt calls burn credits. Don't use the LLM for things a simple Flow lookup or formula can solve deterministically.

• Token limits: Heavy record grounding causes latency spikes. Keep grounding payloads minimal.

5. ALM & Deployments (Where Projects Trip Up)

• Metadata API: Topics, Actions, and GenAiPromptTemplate deploy through standard DevOps pipelines (Copado, Flosum, SFDX).

• Manual Post-Deploy: Playbooks always break here. You must manually:
  1. Assign permissions and FLS to the Agent Execution User.
  2. Activate the Agent version in Target Org.
  3. Validate Data Cloud search index syncs.

This really resonates.

One thing I've noticed is that most benchmarks focus on whether an AI agent can produce the right answer.

In production, that's only one piece of the puzzle.

The bigger questions become: 

• Can it follow business policies? 

• Can it recover from failures? 

• Can it handle long-running workflows? 

• Can you explain every decision it made?

Curious to see how the industry evolves over the next few years.