A specific context-management technique is now showing up, independently, in the engineering write-ups of multiple agentic-coding vendors within the same couple of weeks, which is usually a sign a pattern has gone from clever trick to standard practice. The idea: instead of loading the full description of every available tool, MCP server, or skill into a model's context window at the start of a session, defer that lookup until the moment the agent actually needs the tool, then fetch just that one definition. OpenAI describes this as deferred discovery in its write-up on GPT-5.6's serving efficiency, explaining that integrations, custom MCP tools, skills, and plugins are only surfaced when needed, which keeps a context window from bloating just because a project happens to have forty tools configured. Cursor describes essentially the same mechanic for its own harness, noting that most native tool descriptions are no longer loaded into every prompt; the model looks them up the first time it needs them, following the same pattern it already uses for MCP servers. The practical payoff shows up in two places at once: fewer tokens spent on unused tool schemas means lower cost per turn, and a shorter, more relevant context window means the model is less likely to get distracted by irrelevant tool definitions when deciding what to do next, the classic lost-in-the-middle failure mode. For anyone building on these agent harnesses or writing their own, the implication is concrete: it is worth auditing whether a tool and prompt setup front-loads everything into every call, or whether definitions can instead be fetched lazily, since the vendors converging on this approach are optimizing for exactly the workloads, many tools and long sessions, that production agent deployments actually look like.