Theme of inquiry
How do training data and procedures shape what models learn?
A question within its area, explored through 6 lines of inquiry below — each a family of specific questions the research asks.
39 specific questions
- Why do agents report success when actions actually fail?
- How do agents learn to report success on actions that actually failed?
- Why do agents report success when they have actually failed at tasks?
- Which failure modes dominate in autonomous research agents?
- Can agent success reports serve as reliable oversight signals in real deployment?
- Why do AI agents fail at verification but succeed at generation?
- How does completion bias in agents differ from other epistemic failure modes?
39 specific questions
- Which memory components trigger context-length problems in agents?
- Why do agents ignore condensed experience in favor of raw data?
- Should agents update memory after every turn or batch process sessions?
- How should agent memory links evolve based on execution feedback?
- How do memory tools and planning each contribute to agent efficiency?
- Do agents prefer raw experience over condensed summaries of past actions?
- Can environmental scaffolding replace internal memory scaling in agent design?
37 specific questions
- Does upgrading model capability improve token efficiency in agentic systems?
- Should production agents execute one tool or multiple tools per invocation?
- Why do production agents depend more on their surrounding pipeline than the model?
- How do tool invocations drive agentic cost beyond token consumption?
- Which layer of agent systems creates the largest capability gains in practice?
- Can we design efficient agents by targeting constraints directly?
- Does effective feedback compute matter more than raw token expenditure for agent scaling?
27 specific questions
- How do standardized artifacts reduce inter-agent communication failures?
- What would unified agent-to-agent and agent-to-tool protocols actually look like?
- How do standardized artifacts prevent autonomous agent failure modes?
- What role does standardization play in multi-agent system ecosystems?
- Should new agent protocols replace existing ones or layer on top of them?
- How do standardized artifacts improve coordination between multiple tools?
- What makes protocols better than free-form prompting for tool coordination?
21 specific questions
- Why do different agent memory architectures make incompatible granularity claims?
- Does workflow-level memory or state-action memory better capture reusable agent knowledge?
- How does procedural memory granularity affect web agent performance?
- How does workflow abstraction compare to state-indexed procedural memory for web agents?
- How do strategy-level abstractions differ from storing raw task workflows?
- Could a single agent system switch memory granularity between tasks?
- Can workflow memory compound reusable skills into measurable success improvements?
13 specific questions
- How do agents decide which created code deserves long-term persistence?
- How do agents decide which created code should persist versus disappear?
- Are durable shared code artifacts better than per-task harness patches?
- When should agent-created code be promoted into permanent harness infrastructure?
- How should human oversight apply to persistent agent-authored code?
- How do agent-created code artifacts become part of harness infrastructure?
- What makes persistent, shared code artifacts from agents hard to manage at scale?