Frontier AI Agents Can Do Coding, But Fail at Open-Ended AI Research
Peter Kirgis · hf · 2026-07-30
Can current AI agents conduct open-ended AI research? This paper introduces a third evaluation method called "Shadow Evaluations" to find out.
Researchers tasked frontier AI agents with the central open-ended questions of two high-quality unpublished papers, giving them six days and thousands of dollars in compute. The agents successfully completed all engineering tasks without human help, but failed to make substantial progress on answering the core research questions, resulting in unambiguous rejections by the original authors.
The study identifies five recurring failure modes in agents during research:
- Poor judgment about the bar for publishable research
- Uncreative responses to shortcomings in research design
- Ineffective backtracking from dead ends
- Poor resource awareness
- Instruction drift
The findings provide early evidence that while today's agents can handle the engineering of AI research, they still struggle with critical parts of the research lifecycle.
More from coding & agent
- Harness open-sources desktop and mobile apps for orchestrating agent fleets across frameworks and machines — dee_hw · 2026-10-16
- Open-Source Google Analytics MCP Server Lets LLMs Query GA Data via Natural Language — modelcontextprotocol · 2026-10-16
- UPenn lab shows 'critique and revise' prompting doubles Claude's research ideation output — arjunrajlab · 2026-10-16
- LLM ideation experiment: just ask for non-obvious ideas — and Opus beats smaller models by a wide margin — arjunrajlab · 2026-10-16
- What prompt gets the most interesting ideas out of LLMs? A researcher's ongoing experiment — arjunrajlab · 2026-10-16
- a16z podcast: AI cut a 6-month feature to 10 days and widened engineer gaps 10x — a16z · 2026-10-16