Your AI Agent Is Lying to You and the APIs Are to Blame
Your AI Agent Is Lying to You and the APIs Are to Blame
Your LLM agent called an API, got HTTP 200, and confidently returned a number that was off by 691x. No error. No flag. Just a wrong answer delivered with full authority. If you’re running agents in production against external APIs — and you are — this paper is about your infrastructure, not someone else’s.
What happened
Researchers built SilentGuard, a harness-layer enforcement system that intercepts agentic workflow tool calls and tests each API response against invariants — logical constraints that any correct API result must satisfy — before the result ever reaches the model. The motivation: they found silent failures in 11 of 15 public research APIs tested, with result counts inflated by up to 691 times. The existing mitigations — checklists, reviewer models, LLM-as-judge pipelines — failed badly: agents still reported 77.4% of counts that a judge explicitly flagged as wrong, and those “fixes” left over 70% of answers wrong. SilentGuard moved the decision out of the model entirely, into the harness. Across 7 open models, it cut wrong answers on its target fault types from 74.0% to 1.7% with no regression on clean tasks. On three frontier models, wrong answers fell from 70.2% to 0.5%. The mechanism: when an invariant breaks, the harness withholds the value and abstains rather than attempting recovery. One critical data point — issuing a warning instead of enforcement left 70.1% of answers wrong, meaning soft signals are essentially useless.
Cold read
The 11-of-15 failure rate sounds damning, but we don’t know which APIs, what invariants were written, or how generalizable those invariants are to arbitrary domains — writing correct invariants for your specific APIs is skilled engineering work, not a plug-in. The paper tests SilentGuard on its target fault types to get that 74% → 1.7% improvement; on new fault types the harness alone only reduces wrong answers from 61.1% to 24.5%, which is meaningful but far from solved — and 24.5% wrong is still a production disaster for most use cases. Abstaining rather than recovering is the honest choice, but a system that frequently refuses to answer has its own operational cost: you need to measure abstention rate, not just error rate, or you’ve just traded one failure mode for another. The faithfulness vs. groundedness problem is real, but invariant definition requires domain experts who understand what “correct” looks like — which is exactly what many teams deploying agents don’t have. This is a research prototype evaluated on a curated fault set, not a production-hardened system with known maintenance costs.
What it means for you
- Signal maturity: 3/5 — solid mechanism, genuine numbers, but operationalization gap is wide
- Who gets hurt: Any startup running LLM agents against third-party APIs for data retrieval — analytics tools, research assistants, finance agents, compliance bots — where a confident wrong number is worse than no answer
- What breaks if this is true: Every agentic workflow built on the assumption that HTTP 200 = trustworthy data has silent accuracy rot baked in; your evals are probably also missing it because you’re testing the model, not the tool call layer
- Why it might not land: Invariant authoring doesn’t scale without tooling — most teams can’t enumerate correctness constraints for every API they call, and the paper doesn’t solve that authoring problem
- Watch for: Major agent framework providers (LangChain, LlamaIndex, any Model Context Protocol implementation) shipping built-in invariant enforcement layers; that’s the signal this moved from research to infrastructure
Forecast as of 2026-11-17
By Q3 2027, at least one major agentic framework will ship an optional invariant-checking harness layer influenced by this line of work — but adoption will remain below 20% of production agent deployments because invariant authoring tooling won’t have matured enough to make it low-friction for non-specialists.
Source: Invariant-Aware Harness Enforcement to Protect LLM Agents from Silent API Failures — Sajib Acharjee Dip, Brian Ouyang, Lala Motlhabi. https://arxiv.org/abs/2611.20595v1
