(4) Quick Show of Hands

1. Who used an agentic coding tool this week? (Cursor, Claude Code, Copilot, Codex, …)
2. Who has shipped code that takes action based on LLM output in production?
3. Who has business agents calling APIs autonomously right now?

• If you raised your hand for #1 you've already met one of the two new API consumers

(5) AI Needs Context & Actions

• LLMs alone: stuck in static training data
• Need context (private & current data) and actions (do something)
• Both flow through APIs
• When this layer is missing → GenAI divide

(6) Two New API Consumers

• Coding agents
  • Help developers build integrations faster
  • Consume your APIs at design time
  • AI inference is discovering and coding with APIs
  • The result is traditional code that behaves deterministically
• Business agents
  • Call your APIs autonomously to get things done
  • Read your APIs at runtime
  • AI inference is discovering and using APIs
  • The result are agents that behave stochastically
• Both need things pre-AI developers never required

(7) What Changes for You: DX → AX

• API design: readable for humans (DX) → consumable for machines (AX)
• API abstraction: endpoints → workflows
• Credentials: brokered, so agents connect without direct keys
• Observe & experiment: sandboxes you can learn from
• If your APIs are already in good shape, you're closer than you think

(9) Just Use MCP?

• The pitch: APIs are old, MCP is the future
• Wrap anything in MCP → call it AI-ready
• Skip the hard API work → straight to AI
• Just start fresh after 10 years of API landscape work?

(10) Problem 1: Directly Exposing APIs

• Auto-generate MCP from OpenAPI for every API
• Context window explosion on real workloads
• Bad descriptions → bad MCP
• AI still has to invent workflows across endpoints
• Now your agents get to have a poor integration experience

(11) Problem 2: Investing in the MCP Layer

• Hand-craft MCP servers, design the best tools
• The investment goes into the MCP layer
• The capability layer underneath remains untouched
• Better capabilities benefit everyone, not just MCP
• Everybody's integration experience should benefit

(12) MCP-Washing

• Skip the API work → jump to MCP → call it AI-readiness
• This is the wrong place to invest from a strategy perspective
• MCP should be a thin layer, not a fat one
• Invest in your APIs and then expose them through MCP

(13) Are Your APIs AI-Ready?

• Invest in your APIs
• Start by scoring what you have
• jentic.com/scorecard [https://jentic.com/scorecard]
  • Bad score → fix the basics first
  • Good score → continue

(14) Six Dimensions of AI-Readiness

(16) When LLMs Read Your APIs

• APIs were typically consumed by developers
• Developers write code that often fills DX gaps
• LLMs have more limited context than developers
  • Coding agents shouldn't depend on developers selecting APIs
  • Business agents take the developer out of the picture

(17) Do Your DX Homework

• Most API descriptions need quality improvements anyway
• Better descriptions benefit everyone, not just AI agents
  • Complete and meaningful operation summaries
  • Comprehensive examples for all operations
  • Clear error handling
  • Consistent terminology and naming
• Good DX is the foundation for good AX

(18) Now There's AX Homework, Too

• AI agents need information developers don't typically require:
  • Everything that developers usually just know
  • Treat every API like an external one for developers with no context
• OpenAPI Initiative (OAI) might add AX-specific fields to OpenAPI 3.3
• Until then, leverage description fields strategically

(20) AI Minimalism

• Use AI when you have to, not when you can
• AI execution is orders of magnitude more expensive than running code
• Many business processes are well-defined and repeatable
• Some business processes may need to be auditable and/or explainable
• Your AI strategy should be to use as little AI as possible

(21) What Is Arazzo?

• Declarative: define the data flow and success criteria
• OpenAPI Initiative (OAI) [https://www.openapis.org/] standard for API workflows
• References operationIds from OpenAPI/AsyncAPI documents
• Discoverable: find and run a workflow without knowing its internals

workflows:
  - workflowId: weekend-getaway
    steps:
      - stepId: listHotels
        description: Retrieve hotels available in the destination city
        operationPath: '{$sourceDescriptions.liteapi-hotel-data.url}#/paths/~1data~1hotels/get'
        parameters:
          - name: cityName
            in: query
            value: $inputs.cityName
        successCriteria:
          - condition: $statusCode == 200
        outputs:
          hotelIds: $response.body#/data[*]/hotelId

See the full workflow: arazzo-ui.jentic.com [https://arazzo-ui.jentic.com/?document=https%3A%2F%2Fraw.githubusercontent.com%2Ffrankkilcommins%2Fagentic-tools-webinar%2Frefs%2Fheads%2Fmain%2Farazzo%2Fweekend-getaway.arazzo.yaml]

(23) Today: Agents Get Keys Directly

• Agents are less predictable than traditional developers
• Coding agents produce code → reviewable before it runs
• Business agents produce actions → live as they're decided
• Both need limits, but the controls in place are mostly symbolic

(24) Secure Centralized Access for Agents

• Agents should have their own security regime
• Avoid handing secrets and tokens directly to agents
• Prevent agents from impersonating users or apps
• Ensure proper token delegation and downscoping
• Route all traffic through a centralized access point

(25) Sandboxes for Secure Experimentation

• Create a controlled environment for agents
• Make capabilities available in the sandbox
• Prevent agents from accessing production secrets
• Provide realistic data for testing
• Reduce risk and encourage safe experimentation

(27) Three Levels of API Improvement

1. Improve descriptions — change API description: easy, immediate results
2. Improve definitions — change the API itself: harder, improves capabilities
3. Start platform engineering — built once, used across many APIs:
   • Workflow orchestration (open standard: Arazzo)
   • Credential brokering for safe agent access
   • Sandboxes for safe experimentation

• If your APIs are already in good shape, you're closer than you think

(28) Start Tomorrow: Score Your APIs

jentic.com/scorecard [https://jentic.com/scorecard]

(29) Thank You!

• youtube.com/ErikWilde [https://www.youtube.com/ErikWilde]
• linkedin.com/in/erikwilde [https://www.linkedin.com/in/erikwilde]