Enterprise Agents Go Into Production: Postman Keeps the Keys, Baseten Runs the Engine

Postman and Baseten announced a partnership with a specific goal: get enterprise Agents into production. Postman handles how Agents connect to APIs, how they are governed, and how they run; Baseten handles how models run fast enough, stably enough, and cheaply enough.

Postman’s Agent Stack has four parts. Passport uses scoped credential references so humans and Agents can access APIs without exposing underlying keys. Fabric is an AI-native gateway that routes and governs AI, MCP, and API traffic. Astropods manages runtime, observability, and guardrails in production. Fern generates docs, SDKs, and CLIs so APIs are discoverable, typed, and callable for developers and AI Agents. Postman says it has 40 million developer users and covers 98% of the Fortune 500. That scale means it is more like the entry layer for enterprise Agents.

Baseten positions itself as Postman’s recommended enterprise Agent inference partner. It emphasizes three things: fast, stable, and economical. Model APIs deliver high throughput and low latency; Baseten says seven models released in 2026 ranked in the top two for both throughput and latency in Artificial Analysis, including Beam 1.0, Nemotron 3.5 Super VL, Muse Spark Open, DeepSeek V4.1 Flash 1017, DeepSeek V4.1 Flash, DeepSeek V4 Flash, and GLM-5.2 Fast. On stability, Multi-cloud Capacity Management pools cross-cloud GPU capacity and includes built-in resilience for hardware failures; zero data retention is the default, and customers can restrict or customize cloud, region, and cluster, with support for single-tenant and active-active deployments. On cost, Baseten Training supports post-training open-source models with supervised fine-tuning and reinforcement learning. Harvey used a 27B open-source model for a legal Agent benchmark, matching frontier closed-source performance, while leading closed-source models cost about $50 to run one task.

Postman co-founder and CEO Abhinav Asthana has a line: “If you don’t control your Agent, your Agent will control your company.” In the context of this partnership, it doesn’t feel out of place. Once Agents enter production, credentials, gateways, runtime, and inference costs each turn into bills and incident tickets. Charlie O’Neill, co-head of training at Baseten, also noted that for most economically valuable tasks, there is a big gap between what general models can do through prompting and what specialized models can do through post-training.

The key takeaway from this partnership is the production ledger: the entry point belongs to Postman, the engine belongs to Baseten. Whether an enterprise should do separate post-training for an Agent depends first on whether the task is repetitive enough and expensive enough. If a task costs tens of dollars per call to a frontier model, the math for a 27B specialized model can work out. It doesn’t sound cool, but production environments usually care about exactly this.

Partnership announcement: https://www.baseten.co/blog/baseten-postman-building-agents-for-enterprise/
Original tweet: https://x.com/baseten/status/2115871473516548332