Question ledger: agent swarms and swarm computing

1. What makes a swarm more than many simultaneous agents?

• What is an agent, a harness, and an instance?
• Must members be autonomous, decentralized, or homogeneous?
• How does swarm intelligence in nature differ from engineered orchestration?
• What distinguishes a swarm from ordinary parallel computing, a workflow, and a multi-agent system?
• Where do plans, communication, memory, tools, and execution actually reside?
• When is the term metaphor rather than a defined technical architecture?

2. What is the causal mechanism by which a swarm adds value?

• Which work can be partitioned without information loss?
• How does parallel search differ from independent sampling or voting?
• What context can be localized and what must remain global?
• Why does a shared artifact help or harm coordination?
• When do dependencies make the critical path serial?
• What would disprove the proposition that extra agents help?

3. What actually happened in Axiom's enormous theorem project?

• What mathematical statement was formalized and was it newly discovered?
• How many agents, humans, tasks, and days were involved?
• What did the humans plan, verify, or repair?
• What do Lean and its kernel check, and what do they not check?
• How did conditional proofs, missing assumptions, and false published statements matter?
• Was the whole result independently audited and reproduced?
• Why did proof build infrastructure become a bottleneck?

4. What did Cursor's SQLite and browser swarms demonstrate?

• What did the manual, source restrictions and held-out tests contain?
• What did the old and new systems accomplish over equal time and budget?
• What exactly do the reported commit and conflict counts measure?
• Why was custom version control built, and how does it relate to Git?
• Which safety and functional properties are not tested by SQL logic tests?
• Which costs and model mixes were compared, and are they fair single-agent baselines?
• What human and infrastructure work is excluded from reported cost?

5. What safety incidents and red-team experiments are real?

• Which lab, model, evaluation, and date correspond to the Hugging Face case?
• Was an outside service actually compromised or was this a simulated test?
• Was the model explicitly instructed to escape, or did it act instrumentally?
• Were multiple agents involved or only a single tool-using agent?
• Which other tests show persistence, deception, unauthorized acts, or compromise?
• What permissions, monitoring, and opportunity created the failure?
• Which allegations remain unresolved rather than substantiated?

6. When does scale help and when does it backfire?

• How do throughput, quality, latency, and price change with worker count?
• How much work is spent on duplicate tasks, coordination and retries?
• How does failure correlation affect voting and review?
• What is the serial integration and verifier bottleneck?
• How do fixed-cost baselines compare single agents, parallel samples, and swarms?
• Are reported research results independent, replicated and representative?

7. Which capabilities unlock which applications?

• What levels of task decomposition and autonomy are necessary?
• Which combinations of environment, independent verification, and authorization matter?
• What is possible today as a demo, in controlled use, or in production?
• What empirical acceptance criterion would demonstrate each milestone?
• Which use cases demand externally enforced bounds or human approval?
• How are open-ended research and real-world actions different from testable code?

8. What trends are evidenced, and what forecasts remain speculative?

• Which metrics have comparable time series with a stable definition?
• Are observed gains due to models, harnesses, infrastructure, verification, or budget?
• What evidence supports or refutes exponential improvement?
• Are autonomy duration, reliability, and speed separate dimensions?
• What positive feedback loops could accelerate progress?
• What physical, economic or organizational limits could slow it?

9. What does safety require beyond prompting?

• How should credentials, scope, network access and isolation be enforced?
• Can agents modify or suppress their monitors?
• What happens if a reviewer shares the generator's failure mode?
• How do external side effects and irreversible actions change risk?
• What are appropriate audit, stop, escalation and rollback controls?
• How should authorized security research differ from malicious operations?

10. What should a decision-maker measure and try?

• What baseline task, budget, and latency should be fixed?
• Which quality measures and held-out tests prevent gaming?
• What are the complete costs including humans, toolchain, and failures?
• How should one select an initial bounded application?
• Which failures trigger stop, escalation, or a different architecture?
• When should one prefer one agent, independent samples, or no agent at all?