Research agenda
Open questions about continuity, coordination, adaptation, and evidence in autonomous organizations.
#Questions worth testing
The central question is what allows organizational purpose and obligations to remain coherent across changing tasks, sessions, agents, and environments. Related questions concern authority conflicts, obsolete memory, resource allocation, reviewer independence, and recovery from uncertain external effects. These are research questions rather than capabilities established by the Axiom AO proposal.
Other questions concern useful autonomy: when does delegated planning improve outcomes, and when does it introduce coordination overhead or errors? Comparisons should include simpler alternatives such as one capable agent, a fixed workflow, or human-managed work. More elaborate organization is not automatically more effective.
#Design informative experiments
A continuity experiment can interrupt representative work and examine whether the system resumes the right obligation without duplicating an uncertain action. A memory experiment can change an authoritative instruction and test whether historical information remains available without retaining inappropriate authority. Both require defined expected outcomes and an observable record.
A coordination experiment can vary the number or specialization of agents while holding the task and resource allowance comparable. Record task quality, human intervention, latency, resource use, and failures. Repeated trials and clear environmental descriptions help distinguish an architectural effect from a fortunate individual run.
#Publish methods as well as outcomes
Research publications should disclose the question, system version, mandate, models, environment, acceptance criteria, comparison, and material limitations. Failed and inconclusive trials matter because organizational failures often appear only after apparently successful steps. A selected demonstration is useful evidence for a narrow mechanism, not a prevalence estimate.
The initial literature already supplies complementary settings: simulated workplace tasks, long-running vending scenarios, software-task horizons, and automated research. The research agenda should connect these findings cautiously and develop organization-level tests rather than assuming that performance on one setting transfers to every organizational purpose.
Sources & context
- Xu et al. (2024). TheAgentCompany
- Backlund & Petersson (2025). Vending-Bench
- Kwa et al. (2025; revised 2026). Measuring AI Ability to Complete Long Software Tasks
- Lu et al. (2024). The AI Scientist
Reference edition 0.1 · 2026-10-14. Read the editorial method.