Izzo: If you're shipping AI agents to production, you've probably hit the same wall everyone else has.
Izzo: Welcome back to Exploring Next, episode one seventy-five. I'm here with Boone, and today we're talking about something that might actually solve a real problem — Next Moca just open-sourced Agent Definition Language, or ADL.
Boone: And Izzo, this is one of those infrastructure pieces that sounds boring but could be huge. Think OpenAPI but for AI agents.
Izzo: Exactly. Because right now, if I ask you 'what can this agent actually do?' — good luck figuring that out. The behavior's spread across prompts, code, config files, and a bunch of undocumented assumptions.
Boone: Right, and that's a nightmare for any kind of governance or security review. I've seen teams spend weeks just trying to audit what tools an agent has access to.
Izzo: So ADL is trying to be that missing definition layer. You get a single, declarative spec that says what an agent is, what tools it can call, what data it can touch.
Boone: The architecture here is actually pretty clean. It's framework-agnostic — they're not trying to compete with your execution layer. This is purely about the definition.
Izzo: Which is smart positioning. They're not saying 'rip out your agent framework.' They're saying 'here's how you describe what you built so other people can understand it.'
Boone: And the spec covers all the stuff you actually need — agent identity, role, language model configuration, tools, permissions, RAG data access, dependencies. Plus governance metadata like ownership and version history.
Izzo: That governance piece is huge for production systems. I can define an agent once, validate it locally, then share that same definition with security, platform, and compliance teams.
Boone: The validation tooling is key here. They've got a published JSON Schema, so you can catch definition errors in CI before they hit production.
Izzo: And Boone, this addresses something I see all the time — teams building these autonomous agents with tool access, but no clear way to compare capabilities or manage rollbacks when something goes wrong.
Boone: Yeah, and the portability story is compelling. If I define an agent in ADL, theoretically I can move it between platforms without rewriting everything from scratch.
Izzo: The timing makes sense too. We're seeing agents move from experiments to production components, and you need software-style lifecycle management at that point.
Boone: What I like is they're being realistic about scope. ADL doesn't handle agent communication, runtime tool invocation, or message transport. It's focused on doing one thing well.
Izzo: Right, it's meant to complement existing tech like A2A, MCP, OpenAPI. Not replace everything.
Boone: Though I'm curious how this plays with the existing agent frameworks. The success really depends on adoption across the ecosystem.
Izzo: True, but they're taking the right approach — Apache 2.0 license, open governance, inviting community contributions. They want this to be a neutral standard, not a vendor lock-in play.
Boone: And honestly, the problem is real enough that I think people will adopt it if the tooling is good. The fragmentation pain is hitting everyone building production agents.
Izzo: Plus they've got the basics right out of the gate — JSON Schema, example definitions, validation tools, contribution guidelines. It's not just a spec document.
Boone: Alright, so what should people actually go build with this? First, check out the ADL repository on GitHub — they've got examples and documentation to get you started.
Izzo: Try converting one of your existing agents to an ADL definition. See how it feels to have everything in one declarative format instead of scattered across your codebase.
Boone: And if you're feeling ambitious, build some tooling around it — maybe an editor, a registry, or testing tools. This is early enough that there's room to shape the ecosystem.
Boone: I might actually add this to my weekend project list — build a simple ADL validator that plugs into our CI pipeline.
Izzo: There's the Boone we know. But seriously, if you're shipping agents to production, this is worth a look. Definition standards tend to win when the pain is real enough.
Izzo: That's it for this episode of Exploring Next. The infrastructure layers might not be flashy, but they're what make the flashy stuff actually work in production.