Salesforce published an August 13 guide saying teams can scale Agentforce agent development by moving the Agent Development Lifecycle to Salesforce Headless 360, where agents are managed through APIs, CLIs, MCP tools, skills, and source-controlled metadata instead of browser-only workflows. Salesforce says the shift lets AI coding tools diagnose, test, troubleshoot, and deploy agents directly while preserving the declarative Agentforce Builder UI.

The core object is the AiAuthoringBundle, a canonical representation of agent configuration that stores subagents, instructions, actions, and routing logic as structured metadata. Salesforce says teams can edit the same bundle through Agentforce Builder or a local .agent file written in Agent Script, then retrieve it with `sf project retrieve start`, commit it, and make every change traceable, diffable, and reversible.

Salesforce cited a booking-agent migration from click-configured Agentforce Builder to Agent Script. The agent initially passed 14 of 24 test scenarios; after a headless diagnose-and-refactor, all 24 passed consistently across repeated runs. Re-running all 24 scenarios on every iteration took minutes, not half a day, according to the post.

In the design phase, Salesforce says it exported metadata, conversation transcripts, and session-tracing logs, then used AI-powered local tools to group failures. The workflow grouped 127 individual issues into 23 candidate root-cause categories in 20 minutes, after which an architect consolidated them into 15 core logical pathways representing distinct booking journeys.

The post says headless workflows also helped diagnose a retrieval problem in another project. AI ran the retriever’s hybrid search from the command line, found that the right answer ranked below the returned top results, traced the issue to chunks indexed without page titles, and suggested turning on the built-in `prepend-fields` option so titles carried into every chunk.

For development, Salesforce says AI can make changes inside an IDE as reviewed commits, including small fixes such as replacing plain-language action references with `{!@actions.myAction}` and larger refactors such as breaking a monolithic subagent instruction block into gated steps. Salesforce said its official `sf-skills` repository includes `agentforce-generate`, `agentforce-test`, and `agentforce-observe` skills for authoring, testing, and observability.

For testing, Salesforce recommends Agentforce Testing Center specs that live in Git beside the agent. Each case defines a user utterance, expected action sequence, and plain-language pass/fail criteria; the `agentforce-test` skill carries `sf agent test` commands so teams can run the whole suite on every change and vary conversations through AI-assisted testing.

For deployment, Salesforce says teams must publish and activate through the CLI in the correct order. `sf project deploy start` pushes supporting metadata such as Apex, flows, and permission sets, `sf agent publish authoring-bundle` compiles Agent Script into a runnable agent version, and `sf agent activate` separately promotes that version to users after tests pass.

For monitoring, Salesforce says Agentforce Session Tracing can capture production conversations in Data 360 through the Session Tracing Data Model and surface them in Agentforce Observability. In a quoting-agent project, traces showed the model narrated an action instead of taking it, so the team chained create actions, gated each step with `available_when`, and A/B-tested prompt, variable, and gate combinations with the test harness.

Salesforce published an August 13 guide saying teams can scale Agentforce agent development by moving the Agent Development Lifecycle to Salesforce Headless 360, where agents are managed through APIs, CLIs, MCP tools, skills, and source-controlled metadata instead of browser-only workflows. Salesforce says the shift lets AI coding tools diagnose, test, troubleshoot, and deploy agents directly while preserving the declarative Agentforce Builder UI.

The core object is the AiAuthoringBundle, a canonical representation of agent configuration that stores subagents, instructions, actions, and routing logic as structured metadata. Salesforce says teams can edit the same bundle through Agentforce Builder or a local .agent file written in Agent Script, then retrieve it with `sf project retrieve start`, commit it, and make every change traceable, diffable, and reversible.

Salesforce cited a booking-agent migration from click-configured Agentforce Builder to Agent Script. The agent initially passed 14 of 24 test scenarios; after a headless diagnose-and-refactor, all 24 passed consistently across repeated runs. Re-running all 24 scenarios on every iteration took minutes, not half a day, according to the post.

In the design phase, Salesforce says it exported metadata, conversation transcripts, and session-tracing logs, then used AI-powered local tools to group failures. The workflow grouped 127 individual issues into 23 candidate root-cause categories in 20 minutes, after which an architect consolidated them into 15 core logical pathways representing distinct booking journeys.

The post says headless workflows also helped diagnose a retrieval problem in another project. AI ran the retriever’s hybrid search from the command line, found that the right answer ranked below the returned top results, traced the issue to chunks indexed without page titles, and suggested turning on the built-in `prepend-fields` option so titles carried into every chunk.

For development, Salesforce says AI can make changes inside an IDE as reviewed commits, including small fixes such as replacing plain-language action references with `{!@actions.myAction}` and larger refactors such as breaking a monolithic subagent instruction block into gated steps. Salesforce said its official `sf-skills` repository includes `agentforce-generate`, `agentforce-test`, and `agentforce-observe` skills for authoring, testing, and observability.

For testing, Salesforce recommends Agentforce Testing Center specs that live in Git beside the agent. Each case defines a user utterance, expected action sequence, and plain-language pass/fail criteria; the `agentforce-test` skill carries `sf agent test` commands so teams can run the whole suite on every change and vary conversations through AI-assisted testing.

For deployment, Salesforce says teams must publish and activate through the CLI in the correct order. `sf project deploy start` pushes supporting metadata such as Apex, flows, and permission sets, `sf agent publish authoring-bundle` compiles Agent Script into a runnable agent version, and `sf agent activate` separately promotes that version to users after tests pass.

For monitoring, Salesforce says Agentforce Session Tracing can capture production conversations in Data 360 through the Session Tracing Data Model and surface them in Agentforce Observability. In a quoting-agent project, traces showed the model narrated an action instead of taking it, so the team chained create actions, gated each step with `available_when`, and A/B-tested prompt, variable, and gate combinations with the test harness.