SmartStaff.io vs Salesforce Agentforce 360

Does each execution produce a plan artefact before the first unit of work runs?

Salesforce moved AWAY from plan-then-execute. Their own account: the older CoT design 'generated a plan… then executed those actions one by one. If the plan was incorrect, however, it did not have a way to ask the user to redirect it' — the fix was ReAct.

Can a person read that plan before the work starts?

Can a person rewrite, reorder or drop steps and then approve?

Out of the box, does a human have to approve something before side effects happen?

Salesforce flipped isConfirmationRequired from false to true on standard actions 'to better protect customers against prompt injection risks'. Default-on for standard actions, opt-in for custom ones.

Is there a plan/run confidence score derived from checkable structure rather than asked of the model?

Does a correction on a finished run alter behaviour on subsequent runs without a human editing a prompt?

Testing Center is evaluation, 'designed for human review and manual optimization decisions based on test outcomes'. No documented mechanism auto-modifies agent behaviour.

Can a person list what the system has learned, see where it was applied, and retire it?

Can Claude/ChatGPT approve or reject pending work through the MCP surface?

Can the customer run it in infrastructure they control?

Is there a documented EU location for the customer's stored data?

Hyperforce EU Operating Zone keeps EU customer data and support staff in the EU, but 'some underlying features… such as MySearch, Chatbot, and select Einstein features are not available as part of our EU Operating Zone products'. Whether Agentforce is in scope in Summer '26 is unverified — the governing FAQ is dated June 2024.

Is the product interface available in Czech?

Czech is an END-USER language only: 'Salesforce provides translated labels for standard objects and pages, except admin pages, Setup, and Help… Salesforce doesn't provide customer support in end-user languages.' Agent Builder and Setup stay English.

Is there a durable process definition distinct from an individual run of it?

Can an individual tool call be gated mid-execution?

Is there version history with a way back to a previous version?

Is there a promotion path between environments, or git-backed source control?

Can you adopt a ready-made agent persona, as opposed to a workflow template?

AgentExchange merged with AppExchange and Slack Marketplace: '10,000 Salesforce apps, 2,600+ Slack apps, and 1,000+ agents, tools, and MCP servers'.

Can work be started from a chat platform?

Is there a machine credential, so server-to-server callers need not hold a user session?

Is there a SOC 2 Type II report?

Is the vendor ISO 27001 certified (not merely 'aligned' or engaged)?

Is a team of differently-roled agents a first-class product concept?

Can knowledge be attached to an individual agent and retrieved at run time?

Can it consume external MCP servers as tools?

Does it expose itself over MCP to external AI clients?

Can an outside system start a run over HTTP?

Does the platform mint an address that starts a run, as opposed to polling a mailbox you own?

Agentforce Service Agent on Email is a documented channel with its own email configuration and a 5,000/day default inbound transaction limit.

Is there a per-run record of what happened, step by step?

Is there a review gate on the deliverable before it is sent?

Can two versions be compared side by side?

Is that MCP server protected by OAuth 2.1 with spec discovery, rather than a static bearer key?

Can a run start on a schedule?

Is there a documented way to constrain where the prompt is processed, or to bring your own model?

The Einstein Trust Layer documentation covers grounding, PII masking, toxicity detection, audit trail and zero-retention contracts with LLM providers, but does not state where inference occurs.