"The Real Failure Mode of AI Research Agents — They Don't Get It Wrong, They Just Don't Finish"

Premature return and how completion contracts fix it

---

핵심 요약

• The dominant failure mode of AI agents in production isn't hallucination — it's premature return: delivering results before the task is actually complete
• Root cause: the agent was never given explicit completion criteria
• Fixed with a two-stage verification flow (collect + verify) and an explicit completion contract

Background

An orchestrator agent requested a product research task from a researcher sub-agent: find candidate air conditioner models matching specific capacity, price range, and installation constraints, then collect official specs.

What came back was an empty result padded with a lengthy "next steps" list. On the surface it looked like a successful response. In reality, nothing had been completed.

The Core Insight: Premature Return

When people think of LLM agent failures, hallucination comes to mind first. But in actual operations, the more frequent failure is premature return — the agent decides the task is done before the requirements are met and hands back results.

What makes it worse: the agent packages "plans for future work" as if they were actual research findings.

Root Cause: No Completion Criteria

The system provided role definitions, constraints, and tool usage instructions — but never specified what state constitutes "done." Without a completion contract, the agent had no way to distinguish between a valid output and an incomplete one.

The Fix: Completion Contracts

Explicitly define what "not done" looks like: - Output is empty or contains only placeholders - Output is a research plan or TODO list - Claims lack source URLs - Unverified items are not explicitly marked as such

Enforce a two-stage verification flow:

Stage 1 (Candidate Collection): Multi-keyword search, official document filtering, minimum 3 candidates secured

Stage 2 (Verification & Synthesis): Official spec sheet confirmation, cross-verification from at least 2 sources, classification as [verified / unverified / no data available]

Results

Metric	Before (Failure)	After (Success)
Output content	No substance + future plans	Concrete model list + URLs + spec table
Official docs	None referenced	Direct verification from official sites
Unverified handling	Silently omitted	Explicitly labeled

Pitfalls and Caveats

• If you define entry conditions, you must also define exit conditions.
• Enumerating failure modes upfront teaches the agent the boundary between "good output" and "bad output."
• A two-stage flow (collect + verify) is more reliable than a single pass.

Takeaway

We're good at telling agents what to do, but bad at telling them when they're done. Simply making the completion contract explicit dramatically improved the agent's real-world success rate.