Mastra launches program to build custom AI agent harnesses with companies
Mastra has launched a Custom Harness Program that pairs forward-deployed engineers with company teams to build AI agent systems tailored to their goals, workflows and technology stacks. The program is aimed at organizations that want more control over how agents operate and are measured in production.
- Mastra launched its Custom Harness Program on November 5, 2026.
- Forward-deployed engineers will co-build harnesses in customer repositories and Slack.
- The program includes observability, evaluations, governance and human-in-the-loop workflows.
- Mastra cites customer examples from Salesforce, MongoDB, Workday and Marsh McLennan.
- Workday's move to Mastra took 31 days and cut vector-storage costs by 98%, according to Mastra.
A hands-on program for company-specific agents
Under the program, Mastra's forward-deployed engineers will work inside a customer's repository and Slack while co-building a custom harness. The company says the engagement will cover architecture, observability, evaluations, scoring and workflows designed for production use.
The offering also includes custom MCP, durable human-in-the-loop workflows, tools for teams that will not open an IDE, and authentication, role-based access control and governance features prepared for security review. Customers will also receive early access to the Mastra Enterprise Platform.
Mastra describes the work as outcomes-focused. Its stated goal is to ship an agent system that users adopt and that is designed around the customer's business goals, rather than simply providing a general-purpose development layer.
Why companies are building their own harnesses
Mastra says demand for agent harnesses has grown as models have improved. It points to Salesforce, which in August shipped a harness for all non-Anthropic models to its 100K+ developer community, and MongoDB, which created an internal platform called Sage so every engineering team could ship Mastra agents.
MongoDB's first agent was hand-rolled in Go before the company rebuilt it on Mastra and developed its own plumbing in Sage. According to Mastra, release notes, CI triage and support routing now run on the platform. The company presents this as an example of a harness becoming an internal platform rather than a single application.
The approach can also give companies more control over data and approval processes. Mastra says Marsh McLennan built LenAI on its own infrastructure to keep company data out of consumer chat tools, while Salesforce's harness pauses for human approval before sensitive tool calls.
Examples of cost and workflow gains
Mastra cites Workday's security team as another customer example. The team built a custom harness called SRE to triage alerts to its NOC. Mastra says the agent completes triage in sixty seconds, saving a person-hour every three alerts.
Workday had previously spent months in production on Amazon Bedrock Agents. According to Mastra, moving to the company's open-source platform took 31 days, removed 3,000 lines of code and reduced vector-storage costs by 98%. Workday engineer Bryce Montano said: "After playing around with it for like a week, we were like: Mastra can do everything we need."
MongoDB measures the cost of each agent request against the time saved for a DevOps engineer. Mastra says a few dollars spent on an agent analyzing logs can save an hour of reading, illustrating the program's focus on measurable operational returns.
