Core problem: long-horizon agents can fail quietly without errors, continuing to execute and masking issues.

Google Cloud lays out long-horizon agent challenges and introduces five design patterns to prevent silent, weeks-long failures, with open-source ADK integration and a GitHub code release.

Pattern 3 — Persistent Workspace: ensure the agent’s work state survives across turns by avoiding mid-conversation clears of CLI tooling or session state, and use an execution interface to manage tool state, with production guidance on persistence and sandbox isolation.

Pattern 2 — Background Learning: separate memory extraction from the immediate response; deliver a quick result and update memory in the background, with safety steps like task references, a sibling agent, and throttling to prevent repetition.

Insight: The aim is to detect and fix silent failures early to avoid budget waste or corrupted state, starting with measuring prefix cache hit rates and monitoring prompts that shift every turn.

Pattern 5 — Guard Chain: Normalize credential checks and isolate the model from sensitive endpoints, building a multi-layer guard chain with declarative rules and parsers, plus sandboxing and signed artifacts to prevent exfiltration.

Pattern 1 — Stable Prefix: Move frequently changing prompt content to the end of prompts to boost cache effectiveness, with examples showing prompt size reductions from roughly 70,000 to under 22,000 characters.

Pattern 4 — Explicit Failure: Clearly name terminal states (completed, timeout, halted, pending) to avoid misreporting, and implement safeguards like iteration and tool-call limits per session.

Usage note: These patterns are a Google reference implementation, not a ready-made product; infrastructure considerations such as a managed sandbox for persistent workspaces may be needed, with references to Google Cloud Tech’s X post and ADK samples for details.