Topic
MCP connects agents to tools; A2A connects agents to other agents. Debugging A2A means pinning the incompatible 1.0/0.3 wire version, reading the Agent Card, walking the task lifecycle, streaming events over SSE, and pinning the JWKS out of band.
Read articleYou already have a working REST API. Hand-writing and maintaining a second, parallel set of AI tool wrappers is pure transcription that drifts. Generate MCP tools from your OpenAPI contract, start read-only, and keep authorization where it belongs - in your API.
Read articleYour AI coding assistant keeps using pageSize instead of limit and data.items instead of data.records because it guesses from pasted docs. Give it a local OpenAPI spec it can query over MCP, then verify multi-step scenarios against real requests and mock the missing dependency.
Read articleAn AI-generated schema can be valid, detailed, and wrong. A delegated handler shows why more context is not enough—and what an evidence-based review should check.
Read articleAsking an LLM to read your whole codebase and emit OpenAPI produces fluent, non-reproducible fiction. Pure AST scanning is reproducible but goes blind on dynamic serializers. The reliable design is static analysis first, with AI filling only proven gaps a human can review.
Read articleAn Agent Card is the public resume of an A2A agent: where it lives, which bindings it speaks, and what skills it offers. Fetching it is easy; trusting it is the part teams get wrong. Here is how discovery and signature verification actually work.
Read articleMCP connects one agent to tools and data. A2A connects autonomous agents to each other through discoverable, long-lived tasks. They are not competitors; they sit at different layers, and getting the boundary wrong is where agent architectures become unmaintainable.
Read articleFree-form tool arguments are where agent workflows break: the model invents fields, sends numbers as strings, and omits required values. Structured outputs backed by strict JSON Schema make tool calls deterministic, and most of that schema already exists in your OpenAPI components.
Read articlePointing an agent at forty internal MCP servers breaks tool limits and buries it in duplicated names. The pattern that works at scale is one authenticated gateway that aggregates specs, namespaces the tools, and keeps auth, logging, and rate limits in one place.
Read articleMCP turns external content into actions: a tool description or API response is text the model reads and trusts, which makes indirect prompt injection a real attack path. Here is a practical threat model and the controls that actually work, from scoped tokens to human-in-the-loop and audit logs.
Read articleMost teams expose everything as MCP tools and end up with a catalog the model cannot navigate. The protocol defines three primitives for a reason: tools act, resources are readable context, and prompts package reusable workflows. Here is how to decide, with examples from an API platform.
Read articleAgents discover MCP tools at runtime and learn names and parameters from descriptions, so a renamed tool is a silent production break. Here is how to evolve a tool catalog: additive changes, deprecation windows, versioned names as a last resort, and CI diffs from the source spec.
Read article