What principles should guide the design of agent interactions?

Linear's developer documentation, a "living document" called the Agent Interaction Guidelines (AIG), sets out six principles for "designing agent interactions that integrate more naturally into human workflows." An agent "should always disclose that it's an agent," so it "can never be mistaken for a person." It "should inhabit the platform natively," working "through existing UI patterns and standard actions." It "should provide instant feedback" on invocation, because "silence leads to uncertainty." It "should be clear and transparent about its internal state" — whether "thinking, waiting for input, executing, or finished working" — down to letting a human "inspect the underlying reasoning, tool calls, prompts, and decision logic." It "should respect requests to disengage," stepping back "immediately" and re-engaging "only" on a clear signal. And it "cannot be held accountable": "the final responsibility should always remain with a human."

Linear frames these six as a response to a structural shift, not a style guide: "agents produce work in abundance," so "roles and workflows get reshaped" and "the value shifts to orchestrating input, context engineering, and reviewing output." Each principle maps to a specific failure of that shift — disclosure answers identity confusion when agents and humans "work side by side"; native platform use answers the cost of agents that don't fit existing interaction patterns; instant feedback answers the uncertainty of silence; transparent state answers the opacity of not knowing what an agent is doing; disengagement answers an agent that keeps acting after being told to stop; and non-accountability answers the risk of responsibility diffusing to a system that cannot bear it.

The transparency principle covers the same ground as How can users verify what an AI agent actually plans to do?, which argues that stating goals to an agent is not enough — users need to verify its plans and actions to build common ground. Linear's "inspect the underlying reasoning, tool calls, prompts, and decision logic" names the inspectable surface that verification would act on. Against When should human-agent systems ask for human help?, which proposes six system mechanisms (co-planning, co-tasking, multitasking, action guards, memory, verification) for the same underlying problem of knowing when to defer to a human, Linear's six are pitched one level up, as design principles a platform should satisfy rather than mechanisms a system should implement — "respect requests to disengage" is a principle that an action guard mechanism could satisfy, but the AIG excerpt does not specify how. Compared with Can we design AI autonomy separately from capability?, which structures human-agent interaction by role along an autonomy spectrum, Linear's guidelines instead structure it by behavior — what an agent should always do — independent of how much autonomy it holds.

The excerpt is a developer-facing style guide from one software company, not a study: it offers no evidence that following these six principles improves user outcomes, and the source line marks it "undated" and still evolving, so the principles may change. The excerpt also does not say how the six principles interact when they conflict — for instance, whether "instant feedback" on every invocation can coexist with the "unobtrusive" qualifier Linear also attaches to it, or how "disclose that it's an agent" plays out once an agent is inhabiting the platform natively through UI that looks identical to a human's. The guidelines state a design norm; whether products that adopt it see fewer of the identity-confusion or trust failures the norm targets is not established here.