The recurring architectural bet behind Linear's agent: give the model as little explicit, exhaustive instruction as possible, and instead give it tools to load its own context and goals, then let it reason autonomously in a loop. Over-instructing an agent — spelling out rules, formatting, and edge cases up front — causes it to overfit on unimportant details instead of adapting to the actual situation in front of it. Concretely, this means writing prompts around goals and principles rather than long rule lists, and relying on the model to fetch what it needs (codebase, project description, related tickets) via Dynamic Skills Loading (Agent Meta-Tool Pattern) and tool calls rather than stuffing it all into the prompt. This is the throughline connecting Workflow Entry Points Justify a Native Agent, Two-Tier Model Routing, and Hero-Use-Case Reliability Bar: instruction is minimized everywhere except the small set of hero paths that are deliberately hardened.
An unplanned validation of this principle: users began typing minimal, vague inputs — "linear do the right thing," or even a single emoji — and the agent correctly inferred intent anyway, because it had tool access to load the surrounding context itself rather than depending on an explicit instruction. This behavior surprised the team building it; it wasn't designed for, it fell out of giving the model context-loading tools instead of pre-packaged context.