Agent Deployment Checklist

A single checklist is the artifact that prevents post-deployment incidents. Organized into five blocks: functional tests green, adversarial tests green, observability live, rollback rehearsed, stakeholders signed-off.

Adoption Signals

Every production activation; every material config change (new Invocable, new channel, new persona). Use this skill any time someone asks whether the agent is ready for production.

• Required for any change that adds a new tool to the agent's toolbox or expands record-access scope.
• Run before enabling a new channel (Service Cloud, Slack, Experience Cloud) — channel context changes the prompt-injection threat model.

What Makes Agent Deployment Different

Three platform behaviours, none of which apply to an ordinary Salesforce release, and all of which are documented in Retrieve and Deploy Agent Metadata:

1. A committed agent cannot be edited. "You can't edit a committed agent. Instead, create and edit a new version." There is no quick-fix path — a hotfix costs a full release cycle, so plan the incident response accordingly.
2. Two version counters must be paired by hand. Saves increment AiAuthoringBundle; commits increment Bot/BotVersion. If you save more than you commit they diverge and you must specify both numbers explicitly.
3. Observability is not retroactive. Session Tracing records only conversations occurring after the data model is set up, so it is a pre-activation prerequisite rather than an incident response.

Also: the manifest needs API version 66.0 or later for GenAiPlannerBundle and the new agent metadata types — sourced from The New Agentforce Metadata and Development Lifecycle, not from the Retrieve and Deploy Agent Metadata guide page cited above, whose own example shows <version>65.0</version>. Wildcards on ApexClass/Flows/ GenAiPromptTemplates cause deploy timeouts, deploying a lone BotVersion requires the full agent to be present already, and retrieved agent metadata must not be hand-edited.

Recommended Workflow

1. Functional. sf agent test run --wait green on the routing, golden, and adversarial AiEvaluationDefinition suites. Record the run id on the activation record.
2. Security. Adversarial suite green; PII classification register reviewed against the channel set this activation covers. A channel is part of the approved condition, not a later configuration tweak.
3. Runtime user. Verify the agent user in the target org for Apex class access, object CRUD, field-level access, Flow and Named Credential access, record sharing, and the Data Cloud User permission set. Confirm by invoking each action once as that user — the only check that covers CRUD, FLS, and sharing together. This is the most common cause of "deploy succeeded, agent broken."
4. Observability. Session Tracing and the Session Tracing Data Model enabled before activation, proven with synthetic traffic that appears in the trace data. Alert rules fired at least once and receipt confirmed by the on-call person.
5. Rollback. Rehearsed in a Partial or Full sandbox with a recorded refresh age. Capture the measured duration and the inventory of what did not revert — Apex, custom fields, and activated prompt templates stay at the new version.
6. Sign-off. Business owner, security, and SRE lookups populated on Agent_Activation__c, with a validation rule that blocks save when any gate result is null.

Key Considerations

• The checklist is enforced by a validation rule on the activation record, not by a document. A checklist that cannot block is a memory aid, and memory aids fail under exactly the pressure that makes the rows matter.
• Rollback reverts the agent and nothing that shipped with it. Prefer additive action changes so the rollback surface stays one artefact wide.
• Retaining the prior version is the rollback plan, because a committed version cannot be edited. Retire old versions on a date, not on instinct.
• Sign-off must be queryable, immutable, and linked to a specific version. Chat is none of the three.

Worked Examples (see references/examples.md)

• Rollback rehearsal — Agent v2 activation.
• Stakeholder sign-off record — Quarterly audit.

Common Gotchas (see references/gotchas.md)

• Staging differs from prod — Rehearsal green, prod rollback fails.
• Alert rules not enabled until after go-live — First incident is observed by a customer.
• Sign-off via Slack, no record — Post-mortem cannot reconstruct the decision chain.

Top LLM Anti-Patterns (full list in references/llm-anti-patterns.md)

• Verbal sign-offs.
• Skipping rollback rehearsal because 'the change is small'.
• Dashboards deployed post-activation.

Official Sources Used

• Retrieve and Deploy Agent Metadata — https://developer.salesforce.com/docs/ai/agentforce/guide/agent-dx-deploy-metadata.html
• Agentforce Metadata Types — https://developer.salesforce.com/docs/ai/agentforce/references/agents-metadata-tooling/agents-metadata.html
• Manage an Agent (Agentforce DX) — https://developer.salesforce.com/docs/ai/agentforce/guide/agent-dx-manage.html
• Set Up Agentforce Session Tracing — https://help.salesforce.com/s/articleView?id=ai.generative_ai_session_trace_setup.htm&type=5
• Run Agent Tests (Agentforce DX) — https://developer.salesforce.com/docs/ai/agentforce/guide/agent-dx-test-run.html
• Agentforce Testing Center — https://help.salesforce.com/s/articleView?id=ai.agent_testing_center.htm&type=5