Platform
Dreamforce 2026 for developers: Claude Code, Agent Script, and what to build first
The developer story at Dreamforce 2026 was quieter than the keynote and more usable. An official Claude Code plugin backed by the Salesforce skills library, Agent Script as both the way you write an agent and the way Koa was trained, Apex changes that already shipped in Winter '27, and a trace API you can point at your own observability stack. This is the list, what is available today, and what I would build first.
The keynote was for buyers. The developer announcements at Dreamforce 2026 lived in sessions, release notes and a plugin marketplace, and most of them are things you can install this afternoon.
I have already written the full recap and the admin list. This one is for the person who writes Apex, ships agents from a repo, and wants to know which of last week’s announcements changes how they work. Sorted by what is available now, then by what is dated.
Claude Code is now a supported way to build on Salesforce
The biggest change for a developer is that a coding agent is now an official surface for Salesforce work. The Salesforce Development plugin for Claude Code was announced under the Claudeforce partnership in August, and Salesforce’s developer documentation describes it as a bundle of skills, MCP servers, specialised agents, hooks and commands.
The mechanics matter. When you open Claude Code in a Salesforce DX project, the plugin detects the environment and injects org context. Capabilities resolve in three tiers: a skill first, then the Salesforce CLI, then a Salesforce-hosted MCP server. The plugin ships about 40 skills drawn from the open-source skills library, which holds more than a hundred, covering metadata authoring, Apex, deploy and retrieve with safety gates, security review, reporting, and discovering what is installed in an org versus available.
Salesforce Ben’s Winter ‘27 developer coverage notes the plugin grounds itself through hosted MCP servers, and Salesforce’s developer guide to Dreamforce lists a Metadata Grounding MCP alongside the skills library as the two things that keep a coding agent honest about your org. The same skills library comes pre-installed in Agentforce Vibes.
I covered the Agent Skills format when Salesforce open-sourced it and the DX MCP server before that. What Dreamforce added is the packaging. You no longer assemble the pieces. You install the plugin, authenticate a scratch org, and the safety gates on deploy are the part I would read before anything else.
Agent Script is the source of truth, and now it is training data
Two things happened to Agent Script in the same week. Salesforce’s September 11 announcement describes it as an open-source language combining AI reasoning with deterministic rules. And the Koa paper explains that Salesforce trained its CRM model by expanding Agent Script workflow specifications into scenario- and persona-conditioned tasks, with rewards tied to resolving those tasks correctly.
Read those together. The specification you write to configure an agent is the same artefact Salesforce used to teach a model what a correct CRM action looks like. When Koa reaches general availability, an agent written as explicit Agent Script routes will be the shape it was trained on. An agent that is one topic called “help customers” will not.
That is one more reason to do what I argued for in Agent Script versus prompt engineering: write the routes down. The new Agentforce Builder authors on Agent Script by default, and Agentforce DX lets you keep that spec in a repo with the rest of your metadata.
The Apex changes are already in your org, or will be on October 2
None of the Apex news at Dreamforce was new to anyone who read the Winter ‘27 notes, but the notes arrived in August and the org upgrades land on October 2 and October 9 for most customers. The Winter ‘27 post has the full list. The developer items:
Synchronous heap goes from 6 MB to 10 MB and asynchronous from 12 MB to 25 MB, with a setting to keep the Summer ‘26 limit during a transition. Heap is the limit production code hits least, so treat it as relief for large-payload transforms rather than a design change.
Apex Cursors have been generally available since Spring ‘26 and remain the right tool for read-heavy work over large result sets. The bigger heap widens the slice. It does not remove the need to slice.
Apex integration tests reach developer preview. A class annotated @IntegrationTest can make a real HTTP callout, with @BeforeClass and @TearDown, at the cost of automatic rollback. They run one at a time, asynchronously, behind a scratch-org flag. I wrote up what a real-callout test changes yesterday.
Smaller items worth a look: an Apex Symbol API in beta that exposes type metadata over the Tooling API, which is aimed at exactly the coding-agent grounding above; the latest alias for REST API URIs, which is convenient in tooling and dangerous in anything that processes money; and the beta FORMULA() function in SOQL, which needs API 68.0 and runs in sandboxes, Developer Editions and scratch orgs only.
Traces you can export, and tests you can score
The Agentforce keynote’s best demo was Agent Optimizer reading session traces to find a router gap. For a developer the interesting part is underneath it.
The Session Trace OpenTelemetry API, in beta, returns a full session as OTel spans you can push to Splunk, Datadog, New Relic or any collector. It returns a single session per call and only reaches back 72 hours, so the durable copy is the Session Tracing Data Model in Data 360. I mapped where conversation and trace data lives earlier this month.
On the testing side, Custom Scorers went to beta in Winter ‘27, so you can write the pass or fail question your use case exists to answer as a scorer and run it in the Testing Center. Agent Optimizer arrives in October and will propose configuration changes plus generated tests. A generated test is not independent evidence, so the baseline suite you write now is the thing that makes Optimizer’s proposals reviewable.
The org is now an MCP server, and an A2A peer
Two protocol items change what you can build against.
Salesforce can now host MCP servers that external clients connect to through governed, authenticated connections. That is what Salesforce in Claude, Amazon Quick and the Claude Code plugin all sit on, and it is the mechanism behind the Headless 360 MCP server. The developer consequence is that an invocable action or a Flow you expose today can be called by a client you did not write, as the user who connected it. Design actions as if a model will call them, because one will.
Third-Party Orchestration extends Agentforce across any A2A-compliant ecosystem, with AWS, Azure and Google agents named as delegation targets. The A2A protocol post covers the Agent Card and task lifecycle. The build implication is that the agent you delegate to is one you cannot open in Builder, so the contract at the boundary has to be explicit.
The AIforce framing calls all of this the Headless Toolkit. Strip the name and it is MCP servers, APIs, plug-ins and skills, which is what you were already building against in July.
What I would build first
Five things, in this order.
Install the Claude Code plugin against a scratch org and run its security review skill over Apex you already own. You will learn what the safety gates do before you trust them, and you may find something the Apex security scanner would also have found.
Rewrite one agent in explicit Agent Script routes. Pick the one with the mega-topic. The rewrite is the prerequisite for Optimizer, for Koa, and for any test set worth running.
Point the Session Trace API at your APM in beta and see what a session looks like in the tool your on-call engineers already open.
Write two Custom Scorers for that rewritten agent and put them in a scheduled Testing Center run. That is your baseline.
Try @IntegrationTest in a scratch org against one real endpoint, so you know its constraints before someone asks you to replace the mocks in CI.
Understanding the basics
Is there an official Claude Code plugin for Salesforce?
Yes. The Salesforce Development plugin for Claude Code was announced under the Claudeforce partnership on August 26, 2026. It bundles about 40 skills from Salesforce’s open-source skills library, MCP servers, specialised agents, hooks and commands, detects a DX project automatically, and resolves capabilities through skills first, then the Salesforce CLI, then hosted MCP servers.
What is Agent Script in Agentforce?
Agent Script is the language Agentforce agents are authored in through the new Agentforce Builder. It combines model reasoning with deterministic rules and explicit routes. Salesforce describes it as open source, and the Koa paper shows the same specifications were used to generate the training tasks for Salesforce’s CRM model.
What Apex changes came with Winter ‘27?
Higher heap limits, Apex integration tests with real callouts in developer preview, an Apex Symbol API in beta, a latest alias for REST API versions, and a beta FORMULA() function in SOQL that is limited to non-production orgs. Apex Cursors, which went GA in Spring ‘26, are unchanged.
Can I export Agentforce session traces to Datadog or Splunk?
Yes, in beta. The Session Trace OpenTelemetry API returns one session as OTel spans for any OTLP-compatible collector, but it only reaches back 72 hours and returns one session per request. For history, the Session Tracing Data Model in Data 360 is the durable store.
Write the spec first
Every developer item on this list points the same way. The plugin grounds on your metadata. Optimizer reads your routes. Koa was trained on Agent Script. The trace API shows you what your spec produced at runtime.
The unglamorous work between now and October is writing that spec down properly, one agent at a time. Do that, and each of these tools has something to work with when it arrives.
If you want the Claude Code plugin, Agent Script routes and a scored test baseline set up against your own org before October, talk to us, or run your existing Apex through the Apex security scanner first.
