Can agents ever truly be held accountable for their decisions?

Among the six principles in Linear's Agent Interaction Guidelines (AIG), one states plainly: "An agent cannot be held accountable." The guideline continues: "There should be a clear delegation model between humans and agents. An agent can carry out tasks, but the final responsibility should always remain with a human." Unlike the other five AIG principles, which govern how an agent should behave moment to moment — disclosing itself, giving feedback, signaling state, respecting disengagement — this one is structural: it fixes where responsibility sits regardless of how much of the work the agent actually performs.

Linear's own framing supplies the reasoning: "agents produce work in abundance," which reshapes roles so "the value shifts to orchestrating input, context engineering, and reviewing output." As agents do more of the doing, the guideline's implicit worry is that responsibility could drift toward whoever (or whatever) performed the task rather than whoever authorized it. The fix the AIG proposes is to decouple task performance from accountability by design — "a clear delegation model" — rather than leaving the question to be settled case by case as agents take on more work.

This normative claim bears directly on Can AI systems delegate tasks back to humans?, which describes agentic information systems that can discipline human agents and even delegate tasks back to people — a reversal the paper treats as a structural possibility already latent in agentic IS. Linear's principle reads as a guardrail against exactly that drift: even where an agentic system might functionally delegate to or evaluate a human, the AIG insists the accountability relationship stay fixed in the other direction. Where Linear states the norm — "a clear delegation model" — without specifying what one looks like, How should organizations supervise autonomous AI agents? supplies candidate machinery: digital identity systems, cryptographic receipts for transactions, and immutable logs for every agent action, so organizations prove what agents did, why they made specific decisions, and under whose authority they acted. Deloitte's audit infrastructure is one concrete answer to what Linear's abstract principle calls for. The need for a fixed answer to whose decision was this also echoes Do LLM agents misinterpret delegated requests more than formal protocols?: formal agent-communication languages fix the meaning of a commitment in advance, while Linear's guideline fixes accountability in advance by a different route — a design rule rather than a protocol semantics.

The AIG excerpt states the principle as a design mandate, not as a description of any system that implements it; it does not say what a "clear delegation model" actually consists of, how it would be enforced, or what happens when an agent's output is adopted without clear sign-off from any human. It is also one company's developer-facing guidance, undated and revisable, not a study of whether responsibility in fact stays with humans once agents "produce work in abundance" in practice. The implication the excerpt supports is narrower than the principle's confidence suggests: that accountability should be designed to stay human, not that it reliably does.