The Conway leak was barely old news when Anthropic dropped the real thing. Claude Managed Agents. There’s a list of things that always tagged along when building an agent. Infrastructure, state management, permissions, orchestration. All of it, now bundled into a managed runtime. What took months from prototype to production now takes days, they say (Managed Agents Official Announcement). Exactly the direction the Conway leak predicted. But I didn’t expect it this fast, or this polished. What Shipped A structure that separates sessions, harnesses, and sandboxes. Previously, all three were crammed into one container. A crash wiped session data. Debugging meant accessing a container holding user data. They split it apart. Anthropic’s engineering blog called this “decoupling the brain from the hands,” and the metaphor is accurate (Scaling Managed Agents: Decoupling the brain from the hands). Reasoning (brain) and execution (hands) can now scale independently. And the name is Agents. Plural. Multi-agent orchestration is a built-in assumption. Hand a complex task to one agent and it can spawn sub-agents. Notion, Asana, and Sentry are already running this in production. Pricing is simple. Token costs plus per-session-hour billing. Idle time isn’t charged (Managed Agents Introduction Video). With this kind of structure, the experiment-to-deployment cycle should get quite a bit faster. The Shelf Life of a Harness An interesting point comes up in the engineering blog (Scaling Managed Agents (Engineering Blog)). A harness is essentially “code built on the assumption that the model can’t do X.” When the model improves, that assumption collapses. Case in point: they solved context instability issues with a harness workaround on Sonnet 4.5. On Opus 4.5, the problem simply vanished. That breaks the assumption that harness engineering is a competitive advantage. There was a time when it was. It’s hard to assume that will keep holding. Hence the talk of a “meta-harness.” A structure where internals change but the interface stays stable. But at this point, the thinking shifts. Is being good at writing agent loops itself a durable advantage? If you have to rework the harness every time the model levels up, the ROI on that investment drops. At least right now, the domain side has a better chance of lasting. Why DataNexus Is Safe There’s a problem that keeps showing up. The same word “churn rate” means different things to different teams. Table names like T_CUST_MST are confusing even to humans. An LLM guessing this correctly is even harder. DDL doesn’t carry this context. Taking the definitions that only lived in people’s heads and turning them into an ontology. That’s where it goes. Then feeding it to the model alongside the query. Attach it once and the SQL comes out completely different. But these definitions aren’t publicly available. You can’t scrape it. Take “active customer” alone – marketing says 30 days, CRM says 90 days, and some team defines it as “at least one login.” It’s mostly rules that were set internally. So it’s not the kind of thing a runtime solves. One More Distribution Channel I didn’t see this as a threat. It felt more like gaining one more distribution channel. Think about running DataNexus’s ontology engine on top of Managed Agents – there’s less reason to carry your own infrastructure. MCP wrapping was already in the plan, and it connects naturally with this structure. General-purpose AI is, in the end, general-purpose. To explain why a number looks wrong inside a specific company’s DW, you need the metric formulas and team-specific interpretation rules that live inside. That’s the ontology side. The Wall of Regulated Industries Law, healthcare, accounting, manufacturing. The barriers here are high. Not because of data, but because of the logic inside. Think about hospital data – it’s nearly impossible to construct real logic from externally accessible information alone. A single insurance claim has layers upon layers of internal rules. For a global AI company, digging into all of this doesn’t pencil out. DW/BI is similar. Thousands of tables, abbreviated columns, definitions that differ by team. It’s not just a data problem – it’s the kind where interpretation keeps piling on. Infrastructure Gets Caught Up Fast The speed is what’s interesting. Hours after the Managed Agents announcement, an open-source framework replicating the core functionality appeared (Multica). It always goes like this. One shows up and clones follow fast. Someone builds infrastructure, and almost immediately someone else builds something similar. This layer keeps converging. Doesn’t matter much whose version you use. Direction Check Infrastructure moves up, and what’s left below is data and definitions. Next up is MCP wrapping. Documenting the process of designing and building DataNexus. GitHub | LinkedIn