AI can do a lot, but most AI newsletters leave you wondering what’s actually worth trying. 

That’s where The Shift comes in. They don’t just tell you “what’s new” in AI. 

Every edition answers one question: What can you do with AI today that saves you time, money, or effort?

You’ll see real-world examples, step-by-step mini-guides, and instantly usable prompts under 5 minutes a day.

Plus free access to 3,000+ AI tools and 1000+ mega prompts so you can apply what you learn right away.

No hype, no wasted time, and no “just in case” news. If it’s in The Shift, it’s because it can make your work better today.

And right now, 3 subscribers win a free 1-year Claude Pro subscription. One click to enter.

Subscribe Now - It's Free!

Issue #89

This week's articles explore issues ranging from making existing API estates agent-ready to choosing between SDKs, CLIs and MCP servers as interfaces for agents. We also look at how API platforms are evolving to provide agents with the context and infrastructure they need to operate effectively.

Security and governance feature prominently in this issue, particularly around MCP adoption. We explore emerging approaches to MCP governance, runtime security, authorisation and gateway architecture, alongside some of the vulnerabilities that highlight why these controls matter.

Making APIs accessible to agents in a way that is secure, governed and reliable continues to be the focus API technology vendors.

AI coding agents can connect to repositories, databases, cloud infrastructure and internal APIs through MCP servers, but security teams may have little visibility into which servers developers install or what agents access through them. Snyk has made MCP Governance generally available within its Evo Agentic Development Security platform, introducing a way to monitor and control MCP usage across developer environments.

Axway argues that enterprises do not need to rebuild their API estates to support AI agents. Instead, they need to govern the APIs they already have, make them discoverable and consumable as agent-ready tools, and extend existing security controls to this new class of consumer.

MCP makes it easy for AI agents to connect to tools, but every new server also introduces a potential security and governance risk. Arif Mughal argues that organisations need to treat MCP servers as more than integrations: they are both executable code operating with credentials and sources of information that can influence an agent’s decisions.

AI agents are moving into production, calling APIs, accessing enterprise data and triggering business processes. But Salt Security’s latest research suggests that organisations’ ability to govern these actions is lagging behind adoption: nearly half of surveyed organisations reported or suspected an agent had taken an unintended or unauthorised action in the past year.

API gateways have traditionally focused on exposing APIs to consumers. But as MCP adoption grows, platform teams face a two-sided challenge: connecting agents to external MCP servers while also exposing their own APIs and services as tools.

Giving an AI agent access to an API can mean exposing an SDK, a command-line interface or an MCP server. Each offers a different balance of developer control, flexibility, authentication and context consumption. Postman’s Anthony Viard explores these trade-offs through a real developer workflow, including a surprising cost: MCP tool definitions can consume thousands of tokens before an agent even starts working.

When an API change breaks a downstream service, the problem may not be the code itself, but the dependencies an AI coding agent cannot see. Postman’s new Context Graph API aims to address this blind spot by connecting API specifications, services, repositories, deployments, ownership and runtime relationships into a machine-readable map of the wider API ecosystem.

Giving an AI agent access to an API is easy. Ensuring it chooses the right tool, sends valid arguments and continues working when an external API changes is a much harder problem. Manifest maps 44 tools across four stages of agent reliability: understanding APIs, building integrations, verifying behaviour and running them in production. Its examples reveal how seemingly minor issues, an unsupported parameter or a retired endpoint, can quietly break an otherwise functional agent.

Removing sessions makes MCP simpler to scale, but it also changes how remote servers identify and authorise requests. In its latest article, Curity examines why OAuth access tokens become central to securing stateless MCP, and why the implications extend well beyond the MCP server itself.

MCP security incidents are exposing a recurring weakness: authentication and authorisation controls that appear to be in place but can still be bypassed. Adversa AI’s October roundup brings together recent vulnerability disclosures, attack research and defensive guidance, including an MCP authentication flaw in LiteLLM that reached CISA’s Known Exploited Vulnerabilities catalogue.

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