What’s the secret to staying ahead of the curve in the world of AI? Information. 

Luckily, you can join 2,000,000+ early adopters reading The Rundown AI — the free newsletter that makes you smarter on AI with just a 5-minute read per day.

Sign up to start learning.

Issue #85

Have you seen my podcast episode with Dhayalan from DigitalAPI? If not, it is worth a watch. Here is a clip from the episode.

Turning an existing codebase into a well-documented, agent-ready API can involve a surprising amount of manual work. Elva takes a different approach, discovering APIs within a codebase and turning them into OpenAPI documentation and MCP servers. See the product launch discussion on Product Hunt.

Traditional monitoring can tell you that an API call failed. But when an AI agent fails, the more important question is why it made the decision that led to that failure. Bruno Pedro explores why conventional monitoring and vendor-specific agent observability tools may not be enough for increasingly complex agentic workflows. It makes the case for applying open observability standards such as OpenTelemetry to capture the sequence of decisions, tool calls and outcomes behind an agent’s execution.

Reverb had been running GraphQL for years, but its architecture was becoming increasingly difficult to evolve. The team found a way to modernise incrementally without forcing clients through a disruptive rewrite. The case study offers useful lessons for API teams facing similar legacy architecture challenges.

A major data breach at Manchester Airports Group reportedly involved no sophisticated hack. An API key exposed in public JavaScript gave an attacker access to millions of customer records, raising a much bigger question about how well organisations can actually see and govern the credentials powering their APIs.

Enterprise AI adoption is accelerating, but governance is struggling to keep up. Augusto Marietti analysis highlights a growing gap between the amount of AI traffic organisations are running and the controls they have in place to manage it. The more interesting question is what happens as organisations adopt multiple models and increasingly autonomous agents. Can enterprises govern AI effectively without visibility into the traffic connecting models, APIs and agents?

As enterprises move from AI experimentation to production, governance is becoming less about individual models and more about controlling the APIs, agents and tools connecting the wider AI ecosystem. Cameron McClellan looks at Cisco’s approach to enterprise AI governance and the challenges of bringing greater structure to an increasingly complex AI environment.

OpenAPI has become a standard part of the modern API landscape, but even its name is apparently difficult to standardise. Kin Lane looks at how 62 companies refer to OpenAPI in their own materials, revealing an unexpectedly revealing problem around terminology and consistency. What might seem like a minor naming issue points to a broader challenge: how much consistency can we really expect from an API ecosystem when even foundational terminology varies across organisations?

OpenTelemetry was built to help teams understand what is happening inside their systems. But as adoption grows, organisations are starting to use it for far more than traces, metrics and logs, from billing and infrastructure inventory to AI prompts and application data. That expansion raises an important question: just because OTel can carry the data, should it? Matt LeRay explores where using OpenTelemetry beyond observability makes sense, and where its underlying assumptions could create unexpected problems.

AI is making software development faster, but can API testing keep up? SmartBear put its own API to the test with a new drift-detection capability and uncovered problems that existing tests had missed. The results reveal an important gap between what an API specification says and what the running API actually does, and why that gap becomes harder to spot as development accelerates.

As AI agents become more capable, token usage is becoming an API architecture problem as much as a model-cost problem. Sending unnecessary context through every agent interaction can quietly increase latency, complexity and spend. Kaiwalya Koparkar explores how API and platform teams can reduce this waste by being more deliberate about what information agents actually need, when they need it, and how it is exposed.

Arazzo promises to solve an important problem for AI: describing the workflows behind APIs, rather than simply listing individual endpoints. But Yoan Gross argues that while the idea is right, the specification may not yet provide everything AI tools need. The article examines where Arazzo fits in an increasingly agent-driven API ecosystem and introduces Bump.sh's Flower Spec, which builds on Arazzo.

The API Platforms For Scale Newsletter curates intelligence on the trends, technologies, and practices shaping API platforms.