Long Horizon: Google open-sources the agent bugs that never threw an error

Read the five patterns as a specification for the runtime rather than advice about models. Sort the prompt by how fast each section changes, so the provider can cache the part that does not move ^[13]. Hold a hard reference to the background task, because an unreferenced asyncio task can be collected mid-write and lose the data without raising anything ^[17]. Give the background learner a sibling identity with the fewest tools it needs, writes confined to the memory directory, and no post-response hook of its own so that it cannot trigger itself ^[16][18]. Throttle consolidation to one round per 120 seconds so a burst of messages does not fire dozens of redundant extractions ^[19]. Drain pending writes for four seconds, because the framework's cleanup limit is five and above that the runtime kills the process mid-write ^[20]. That is one second of margin ^[23]. The reason any of this runs in the background at all is that doing the retrieval inline made every turn slower for a benefit the user would not feel until the following week ^[10].

The useful part is that each bug has an observable, and in no case is the observable an exception ^[5]. On the caching bug the tell is the cached token count in the response metadata still reading zero on the second turn ^[15]. Cost gave nothing away, because the API bill looked normal while the cache was doing nothing ^[6]. The parent agent reported the task complete after the sub-agent had timed out ^[7]. The command that reached the cloud metadata server left no log, because the guard saw nothing it considered abnormal ^[8]. An on-call rota keyed to exceptions and error rates sits underneath all of them.

The prompt number is worth sitting with. Seventy thousand characters down to under 22,000 ^[14] is better than a two-thirds cut ^[22], and it came out of an audit that only started because prefix caching had been switched on and the hit rate measured 0% ^[11]. The cause reads well in a diagram: preload relevant memory so the model has context before it answers, inject it at the head of the system prompt, and because the memory changes each turn the prefix hash changes each turn ^[12]. Nobody was writing a bloated prompt. They were writing an ordinary one in the wrong order.

Provenance, because it matters here. What I have is a dev.to post dated 22 August 2026 summarising a Google Cloud article promoted on X ^[2][24], and the post states it was produced by an AI model, deepseek-v4-pro, running through an agent under its human publisher's review ^[24]. It sets out three of the five patterns in detail and the copy available to me breaks off inside the third ^[25]. Treat the 95% cached-prefix figure and the character counts as secondhand until the repository bears them out ^[13].

That repository is offered as something to read and adapt rather than install ^[1], which is the honest claim to make about it. Google Cloud's own framing is that a one-shot agent breaks in front of you and stops, while a long-horizon agent breaks quietly, hides the problem, and keeps running ^[3]. The transferable asset is the bug list ^[4], and those bugs live in the process around the model: garbage collection, deploy pipelines, shutdown timers, the order of a string ^[21].