Issue #88

AI will write 90% of code by the end of 2026, and only 10% of developers will stay relevant.

The engineers in that 10% aren't smarter. They just know how to use AI.

We put together the exact playbook Anthropic engineers use to make sure you're in the top 10%.

Sign up for The Code and get access to:

Join 250K+ engineers

API-first organisations, true to their name, provide APIs. But many are on the journey to becoming AI-native and, as part of that transition, thinking agent-first. This means providing agentic capabilities such as MCP servers and agent skills.

But how do you package these capabilities to ensure they not only improve the agentic experience but also deliver measurable business impact?

There are three things I would say about this.

1. Package Agentic Capabilities as Agent Plugins

It is important to package MCP servers and agent skills as agent plugins. Agent plugins provide a vendor-neutral way of bundling skills, MCP servers and other agent extensions. Essentially, an agent plugin is a directory with a standard format that solves a key problem: it serves as a wrapper that enables different clients to discover and use your skills and MCP servers.

As a bundle of agentic capabilities, an agent plugin also makes it easier to distribute a capability bundle across agent plugin marketplaces, GitHub and developer portals.

2. Treat Agent Plugins as Products

The second point is that agent plugins need to be managed as products. That means understanding the intended human user, their use case and how agents can help them fulfil it. It requires clear, measurable outcome metrics for the human experience, as well as metrics for the agentic experience, for example, time taken, tokens consumed and reduced errors. This would require the full discipline of product management.

3. Design Agent Plugins Deliberately

Third, creating agent plugins involves a number of software design decisions:

In conclusion, organisations need to distribute agentic capabilities through agent plugins, but these plugins need to be treated with the same disciplined product management approach as any other digital product.

On that note, here are some interesting links for you this week.

As AI agents spread across Uber, teams began building their own MCP integrations, creating duplicated infrastructure, fragmented tooling and growing security challenges. Uber explains how it combines API discovery, protocol translation, centralised governance and runtime security, while addressing two less obvious challenges: helping agents find the right tools and reducing the token costs of exposing thousands of them.

Not every capability needs to be a fully autonomous agent. Tommaso Stocchi explores what happens when specialist agents are replaced with distributed Agent Skills and MCP tools, allowing the main agent to retain the reasoning while remote services provide domain-specific instructions and operations.

As AI applications and agents increasingly interact with models, APIs and tools, traditional API governance needs to expand. API7’s guide argues that organisations need to govern not just AI traffic, but the identities, models, providers, agents and MCP tools behind it.

The next phase of AI may be shaped less by smarter models and more by standard interfaces that allow agents, tools and other agents to work together. Tommy Tran examines how MCP, A2A, Agent Skills and emerging observability standards are beginning to create a common infrastructure for agent interoperability.

Building an AI agent once meant writing the loop yourself: call the model, invoke tools, manage state and repeat. This article argues that model is changing as execution, identity, tools, observability, workflows and policy move into managed runtimes.

APIs were largely designed around human-driven applications. AI agents introduce a very different kind of consumer, one that can discover tools, make decisions and execute API calls autonomously and at scale.Art Anthony explores what that shift means for API identity, authorisation, security, governance and observability.

Kong’s latest AI Summit announcements point to a broader shift: AI infrastructure is moving beyond model access towards connecting and governing agents, tools, APIs and data as one ecosystem.

The conversation around AI agents is moving beyond models and frameworks towards the infrastructure needed to operate agents reliably in production. Following AGNTCon and MCPCon in Amsterdam, Matt Deutsch highlights the growing need for security, identity, governance, observability and runtime management as agents move out of experimentation.

As enterprises accumulate API gateways, identity platforms, integration tools and AI infrastructure, technology fragmentation is becoming a governance problem as much as a cost problem. WSO2’s guide explores why CIOs are increasingly looking towards platform consolidation to bring these capabilities under a more unified control model.

Making an API “agent-ready” may require more than exposing endpoints through MCP. Daniel Kocot argues that agents are a new type of API consumer, with different needs around discovery, guidance, planning, execution and understanding the current state of a resource.

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