A context reset is a few lines of harness logic, and deleting it costs nothing. What expired was the belief underneath it: that Claude needed help finishing work before it ran out of context [1][2]. Anthropic's engineering account puts that mechanism plainly, saying harnesses encode assumptions about what the model cannot do on its own and that those assumptions go stale as models improve [1]. One dead flag is not a write-off in isolation. As accounting that holds; as a maintenance plan it does not, because a harness is a stack of beliefs about model deficits and no release note tells you which of them just became false.
The container story in the same account is the costlier lesson, and it has nothing to do with model quality. Anthropic began with session, harness and sandbox in one container, which bought direct syscalls for file edits and no service boundaries to design [7]. The bill arrived in diagnosis. The only window into a stuck session was the WebSocket event stream, and three different failures - a harness bug, a dropped packet in the stream, a container going offline - presented identically through it [9][16].
The durable artifact Anthropic now proposes is the set of boundaries around the harness: a session as an append-only log, a harness as the loop that calls Claude and routes tool calls, a sandbox as the place code runs, each swappable without disturbing the others [6], and each able to fail or be replaced on its own [12]. The justification offered is the operating system precedent, where process and file abstractions outlasted the hardware beneath them and read() no longer cares whether it is addressing a 1970s disk pack or an SSD [5]. That is an argument by analogy, and the record contains no history of these particular interfaces surviving a model generation. Anthropic says it expects harnesses to keep evolving [15] and that its interfaces are meant to outlast any implementation, including the ones it runs today [4] - a claim the company has not yet had the chance to test.
So the trade is legible even where the evidence is thin. A team that keeps building its own harness keeps paying to re-verify assumptions it cannot see listed anywhere, and keeps owning the debugging problem that coupling creates. A team that moves to a hosted service sheds that rework and takes on a boundary defined by the same vendor whose release cadence produces the depreciation. Anthropic is explicit that it is opinionated about the shape of the interfaces and not about what runs behind them [14], which is the part of the offer a buyer can actually test over time.
The decision available this quarter is narrower than the decade-long claim about abstractions. It is whether a session survives the death of the machine it ran on, and whether an engineer can diagnose a stuck agent without opening a shell in a container that holds customer data [10]. Answer those two and the next model release removes a workaround without removing your ability to ship; leave them coupled and the release that deletes your scaffolding also deletes your ability to tell what broke.
