Claim
When an AI agent loses track mid-task, expanding the context window does not solve it. The real ceiling is memory architecture: agents that load everything into context degrade, while agents that retrieve only the relevant skills at task start stay sharp.
Mechanism
Tunguz describes a three-layer system he calls the preflight check, borrowed from aviation pre-departure planning. Before executing any task, the system: (1) matches the incoming query to a library of stored skills via intent matching; (2) loads only the matched skills into the active context window; (3) executes the task locally for routine work, routing to frontier models for complexity. A background watchdog monitors overnight, logging which skills fired and which decisions succeeded. Successful patterns harden into deterministic code, reducing LLM reliance for stable work. The self-improving loop means the system gets cheaper and more reliable over time without any explicit retraining.
Conditions
Holds when: the agent operates across many task types and needs to handle routine workloads efficiently. Gains the most when the skills library covers 70-80% of tasks.
Fails when: every task is novel with no repeatable pattern to load. Pure research agents or frontier-only creative tasks don't benefit from a preflight retrieval step.
Evidence
Tunguz writes: "Preflight is retrieval. The agent inspects its skills library, picks the ones relevant to the task, and loads only those into the context window." He adds: "Context size is not the ceiling. Memory architecture is."
Signals
- Agent performance stays consistent across long conversations.
- Routine task costs fall as more work shifts to deterministic code paths.
- The skills library grows over weeks without developer intervention.
Counter-evidence
Maintaining a separate skills library adds system complexity and requires a curation layer. For teams without a watchdog process, the library can stagnate and diverge from current task patterns. The architecture is more complex to debug than a single-context approach.
Cross-references
- Related to Routing architecture is the first design decision in an AI agent system, not model selection: routing by task complexity (async vs. synchronous) is complementary to preflight skill retrieval.
- Related to Two-tier local/cloud AI routing cuts response latency by 60% and lifts throughput 25% by keeping simple tasks on-device: the local-first execution strategy applies to the preflight's 70-80% routine workload tier.