Anthropic’s agentic push forces competitors to add memory and Advisor‑style strategies

Anthropic is pushing a new way to build AI agents: keep a cheaper model in charge, and call a stronger one only when the task gets hard. (claude.com) On April 9, Anthropic introduced an “advisor tool” on the Claude Platform. The setup lets Claude Sonnet or Haiku act as the executor while Claude Opus serves as the advisor inside the same Messages API request. (claude.com) Anthropic said the executor handles the full task loop — calling tools, reading results, and deciding next steps — and consults Opus only when it “can’t reasonably solve” a decision on its own. The advisor returns guidance, a correction, or a stop signal, then the executor continues. (claude.com) The company’s pitch is cost control. Anthropic said Sonnet with an Opus advisor improved by 2.7 percentage points on SWE-bench Multilingual versus Sonnet alone, while reducing cost per agentic task by 11.9%. (claude.com) Anthropic also published a bigger jump for lower-cost setups. On BrowseComp, Haiku with an Opus advisor scored 41.2%, more than double Haiku alone at 19.7%, while still costing 85% less per task than Sonnet solo. (claude.com) This fits a broader Anthropic product push around longer-running, context-heavy work. In November 2025, the company launched memory for team users, saying Claude can remember project details and preferences across conversations with project-specific boundaries and user controls. (anthropic.com) Two weeks ago, Anthropic’s engineering team described Managed Agents as a hosted service for “long-horizon agent work.” The company said it built stable interfaces for sessions, harnesses, and sandboxes so developers would not have to keep rewriting orchestration as models improve. (anthropic.com) That combination — memory, hosted agent infrastructure, and model-to-model escalation — moves Anthropic further beyond selling a single chatbot or application programming interface endpoint. It gives developers more of the plumbing needed to ship domain-specific agents for coding, research, and enterprise workflows. (anthropic.com 1) (anthropic.com 2) (claude.com) Anthropic reinforced that agent focus again on April 16 with Claude Opus 4.7, which it said was optimized for “complex, long-running tasks” and was available across Claude products, the API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry at the same price as Opus 4.6. (anthropic.com) The practical effect for startups is that “wrapping” one frontier model with a thin interface looks less differentiated when the model vendor is shipping memory, routing logic, and hosted agents itself. The value shifts toward vertical products, proprietary workflows, and customer data that the base model company does not already provide. (claude.com) (anthropic.com 1) (anthropic.com 2) Anthropic’s latest releases do not settle the broader platform race, but they make the battleground clearer: less about a single best model, and more about who owns the memory, the routing, and the agent loop around it. (claude.com) (anthropic.com)