A working checklist for agents that have to survive longer than one call

n8n's Andrew Green argues that prompt, context and loop engineering all still leave the model responsible for correctness, and sets out what to build instead. Treat context as a lifecycle artefact: watch the token split across system prompt, tools and history, compress older events with sliding windows, and write to an append-only ledger with identity-scoped retrieval and TTL expiry. Make execution durable by persisting agent state, scheduled tasks and connection state while letting compute sleep, and waking on webhook callbacks or backed-off polling, with an append-only log that replays deterministically after a crash. Above all, replace the model-as-judge with deterministic gates — status codes, schema validation, cross-field consistency, state-diff re-queries, finite state machines, and small encoder-only classifiers that return a verdict rather than prose. The named failure modes are context rot, hallucination multiplication from LLM-judging-LLM, mid-JSON truncation corrupting tool calls, recursive tool loops, and duplicate side effects from non-idempotent webhook retries.

What it means The single line worth stealing: write down the done condition before the agent starts, and check it with something deterministic rather than another model.