Copado, Gearset and Provar AI agents: a client check

Copado Agentia, Gearset Cam and Provar MCP are the 2026 AI agents of three Salesforce DevOps and testing vendors, and everything about them rests on vendor claims. Copado says Agentia is available in its Advanced and Pro editions, Gearset launched Cam with pilot customers, and Provar MCP is a free local plugin for Provar Automation customers. Before a manufacturer pays, I would ask four questions: which model runs, where the data is processed, how long it is kept and who approves what the agent produces. Of the two selling their own agent, only Gearset answers part of that publicly.

What do Copado Agentia, Gearset Cam and Provar MCP do?

The descriptions below are the vendors’ own; I found no independent test on Salesforce Ben or comparable Salesforce press.

• Copado Agentia. Announced on 13 April 2026 in Copado’s press release1. A Context Hub combines org metadata, dependencies, pipelines and history with your own documentation and policies. On top sit agents for plan, build, test, release and operate, an Orchestrate Agent that coordinates them, and Agentia Studio for custom workflows and Agentforce agents. Built-in governance, per Copado: policy controls, automated testing, approval gates, audit trails and role-based access.
• Gearset Cam. According to Gearset’s announcement of 2 September 20262, Cam takes a ticket, asks for clarification when the request is unclear, plans and builds the change and checks that it is deployable before it opens a pull request. Its work runs through the team’s release process with the same tests and reviews as human changes, with an audit trail.
• Provar MCP. Published on 1 July 2026 on the Provar blog3. This open-source Salesforce CLI plugin lets an AI assistant such as Claude, GitHub Copilot or Cursor read org metadata, generate Provar tests and Page Objects, validate them against more than 240 quality rules, run them and triage failures.

Copado and Gearset sell an agent that changes your org. Provar adds a tool to the AI assistant your developers already use, which makes its due diligence much shorter.

What is generally available and what is still a pilot

A pilot rarely comes with a service level or a fixed price, so status matters. As of 27 October 2026:

Product	Status as the vendor states it	Price published	Model named
Copado Agentia	“available today” in the Agentia Advanced and Agentia Pro editions (April 2026)	no	no
Copado Agentia Headless (agents in the IDE)	“officially GA” on Copado’s Dreamforce 2026 page, no date given	no	no
Gearset Cam	launched with pilots at “a selection of Gearset customers”, no GA label	no	no
Provar MCP	included with a Provar Automation licence, CLI plugin installed from the beta tag	no extra cost	uses your AI assistant

Sources: Copado press release, Copado Dreamforce 2026 page4, Gearset announcement and Gearset agent page5, Provar MCP documentation6.

The one hard number comes from Gearset:

44%

of a team's pull requests were created by Cam at one customer within a week

75%

of these pull requests were successfully merged

That is one customer, measured by the vendor, after one week. It shows that Cam can produce mergeable pull requests in a real org, nothing more. One in four was not merged, and every one went through the team’s review.

Model, region and retention set the three apart most

Here the three products differ most, and this is where a manufacturer with GDPR obligations and a works council should start.

Gearset publishes general AI terms in its product and location specific terms7, in a version effective from 15 September 2026. They do not mention Cam by name, but they cover Gearset’s AI functionality as a whole:

• Providers include “OpenAI and Amazon Web Services”, with “data centres in the United States and as otherwise specified in the Documentation”.
• Gearset keeps input and output for 90 days or, where the software offers it, a shorter period you set; the providers keep nothing.
• Gearset does not train models on customer data, but may use input and output to assess and improve its services.
• Usage is measured in credits with a monthly cap from the order form. At the cap, Gearset suspends the AI functionality until the end of the month.
• The customer “may not use the AI functionality to replace or substantially replace human decision making”.

Copado names no model and no region for Agentia on the pages I opened. Its public data sheet “Copado AI and your data” dates from June 2024 and covers CopadoGPT and Test Copilot, not Agentia; I could read only its summary page. That says nothing about the quality of the answers; it means you need them in writing before you sign.

Provar MCP runs “entirely on your local machine”, according to its documentation, and sends “no project files, test code, or credentials” to Provar servers unless you switch on an optional API validation. The metadata and test cases still reach whichever AI assistant you connect. The data question moves to your contract with Anthropic, GitHub or Cursor; the works council and GDPR article covers that part. As I read the vendors’ pages, Provar MCP is the easiest of the three to approve. My reading of these terms is not legal advice.

Worked example: a machine builder with Copado and a nearshore partner

Assume a machine builder with Sales Cloud for dealer sales, Field Service for its service technicians and an SAP S/4HANA integration for spare parts and warranty. The pipeline runs on Copado. Three internal developers work with a partner who provides six nearshore developers, and the partner already codes with Claude Code through accounts the client owns. All numbers here are assumptions.

Copado offers an upgrade to Agentia. This is how I would handle it:

1. Separate building from governing. The coding agent already writes Apex and Flows. The question for Agentia is whether its Context Hub and approval gates add something the existing pipeline and the quality gates for AI code do not.
2. Get the four answers in writing. Model and provider, processing region, retention, and how usage is metered and capped. Until they are on paper, Agentia only sees a developer sandbox without customer or dealer data.
3. Pilot one narrow ticket type. For example 20 small changes on the warranty claim object: new fields, picklist values, a validation rule. Measure merges without rework and review time against similar tickets done by the coding agent alone.
4. Add Provar MCP if Provar is already licensed. The regression tests for the dealer portal are the obvious target. It costs nothing extra, and the data stays inside a tool chain that is already approved.
5. Decide after six weeks on numbers from your own org, not on the vendor’s 44 and 75 percent.

With Gearset the steps are the same; step 2 starts from Gearset’s published terms and asks for an EU region.

When is open source plus a general coding agent enough?

A team that already runs sfdx-hardis gets much of what paid DevOps agents offer without a second AI contract. According to the sfdx-hardis AI setup guide8, it explains deployment errors and describes Flows and Apex classes. The provider can be Agentforce, OpenAI, Anthropic, Google or a local Ollama model, set by environment variables or routed through a corporate gateway. You choose the model and therefore the region. Its coding agent auto-fix is in beta, and the guide says an expert must review those changes before merging.

Pair that with Claude Code, Codex or Copilot, compared in Agentforce Vibes, Claude Code, Codex or Copilot, and a pipeline you own decides what reaches production. The trade-off is support.

Put these questions in writing before you pay

I would put these questions to any DevOps vendor selling an AI agent and get the answers into the order form:

• Which model and which provider run the agent, and will you tell us before you change them?
• In which region is input processed, and is an EU region available?
• How long are prompts, code and metadata kept, by you and by the model provider, and are they used for training?
• How is usage metered, what is the monthly cap, and what happens when we reach it?
• Is the product generally available, in beta or in a pilot, and which service level applies?
• Who approves the agent’s pull requests, and can it ever deploy without a human?
• What happens to the context the agent built about our org when we leave?

The last two follow from two rules: the client should own the DevOps tools, and the same terms belong in partner contracts for AI coding. Setting this up is part of my DevOps and release management work; the wider tool choice is in chapter 4 of the AI masterclass.

Sources8

1. 1Copado’s press release (opens in a new tab)copado.com
2. 2Gearset’s announcement of 2 September 2026 (opens in a new tab)assets.gearset.com
3. 3Provar blog (opens in a new tab)provar.com
4. 4Copado Dreamforce 2026 page (opens in a new tab)copado.com
5. 5Gearset agent page (opens in a new tab)gearset.com
6. 6Provar MCP documentation (opens in a new tab)documentation.provar.com
7. 7product and location specific terms (opens in a new tab)gearset.com
8. 8sfdx-hardis AI setup guide (opens in a new tab)sfdx-hardis.cloudity.com