The Pulse
Eon Benchmark Finds Enterprise Agents Miss Hidden Facts
Eon researchers tested 12 enterprise-agent setups on questions that required inferring facts across simulated company records. The best setup answered 18 of 24 attempts correctly, while similar-record questions stumped nearly every agent.
AI.info Team ·
Enterprise agents that scored 22 to 25 out of 27 on straightforward company-data questions struggled when the answer depended on clues scattered across records. A paper submitted to arXiv on September 24 introduces eight such tests for the Era by Eon benchmark, measuring whether agents can infer facts that no single document states. In the authors’ test, the top-performing setup got 18 of 24 attempts right.
From stated rules to hidden facts
Era by Eon generates a fictional company and serves its records through simulated business applications. Its earlier questions spelled out the rules needed to calculate an answer, allowing code-enabled agents to search and compute. The new questions instead require agents to connect clues across systems, such as a sales record, a customer call and a support ticket.
One example centers on a lost sale. The sales system attributes the customer’s decision to timing, but a recorded call points to a service outage; identifying the actual cause requires finding the relevant call and linking it to the right support ticket. The benchmark’s eight templates also test questions about service credits, departed contacts, conflicting deal amounts, roadmap work, signed renewal offers and how to reconstruct a sales pipeline from an old report.
Claude Fable 5.1 leads the test
The researchers tested 12 configurations, pairing six language models with two agent programs. One program could not run code; the LangGraph-based setup could run scripts in a sandbox. Each configuration attempted each of the eight questions three times, for 24 runs.
Claude Fable 5.1 scored 18 correct runs with the non-coding Era agent program and 15 with LangGraph. GPT-6 Astra scored 13 and 12, respectively. The other four models scored no more than six correct runs with either program, according to the paper.
Code made a large difference on the benchmark’s computable questions: GPT-6 Astra answered 24 of 27 with LangGraph, compared with eight using the Era program. It did not produce the same benefit on the hidden-fact test. For Claude Fable 5.1, the score was higher without code than with it.
Similar records expose a different weakness
The hardest questions asked agents to identify the one relevant record among several plausible candidates. In one example, a customer received three renewal offers, and the agent had to determine which offer the customer signed. On two such questions, the twelve agents together answered correctly in 1 of 84 attempts.
That gap matters for workplace use because company knowledge often sits across separate systems and conversations, not in a single field. Yet the results come from eight questions about one generated company, not a broad sample of real deployments. The authors argue that what the agents lack is not much:
“The missing knowledge is small. It is a handful of facts of the kind that an employee learns in the first weeks: which system holds the committed amount, on which day the roadmap moved, how the pipeline review is computed.”
The authors of the Era by Eon paper
A written store of such facts, they write, would turn the hidden facts into recorded ones, and the benchmark could measure how much it helps by running the same agents with and without it. The paper does not report that experiment.