September 2026 Summaries

Arazzo, an OpenAPI Initiative specification for documenting multi-step API workflows, has gained renewed relevance as AI agents increasingly orchestrate APIs through MCP servers, tools, CLIs, and direct calls, where unreliable API-chain inference can cause hallucinations, latency, and high token costs. Although Arazzo was initially less compelling for human-facing API documentation and explorers, it became more useful for agentic execution, yet the author identifies limitations in its two-layer conditional logic, mixed runtime-expression syntaxes, inability to transform filtered outputs effectively, lack of current human-in-the-loop elicitation support, and potential drift caused by separate OpenAPI and workflow files. To address these gaps, the author’s team created Flower, an open-source internal workflow model that can import Arazzo documents and offers simplified step conditions, dot notation with JMESPath-based transformations, and self-contained workflows without OpenAPI dependencies. Despite Flower’s growing use among clients, the author views it as an interim alternative and hopes future Arazzo versions mature into a broadly adopted standard for reliable AI-driven API workflows.

Companies often duplicate existing product capabilities when building embedded AI agents and later MCP servers, creating parallel integrations that can drift as APIs, permissions, schemas, error handling, and security requirements evolve. The proposed approach treats APIs, workflows, MCP servers, and agents as layers of one stack: the product backend remains the source of truth for business rules and side effects, APIs provide programmable contracts, workflows translate user intents into deterministic API sequences, MCP standardizes tool discovery and execution for AI clients, and agents focus on conversation and user experience. Rather than manually maintaining separate integrations, organizations can derive and validate agent-facing capabilities from API contracts, using established API lifecycle practices such as versioning, testing, CI/CD, and breaking-change detection. Citing Postman’s 2025 report that 82% of organizations have adopted API-first practices but only 24% design APIs for agents, the discussion argues that existing API infrastructure should be reused as agent adoption grows. Bump.sh presents its MCP Platform as an implementation of this model, enabling teams to define API-based workflows with Arazzo or Flower and deploy them as hosted MCP servers with authentication, secrets management, execution, and observability for multiple AI interfaces.