Open research question
Question & scope
We study the architectures, protocols and operating conditions that allow AI Agents to plan, use tools, collaborate and recover safely across long-running tasks.
This page keeps the question open while making it researchable. It separates the object being evaluated, the operating context, the evidence needed to support a claim and the conditions that could disprove it.
A reader should leave with a defined system boundary, a usable evidence plan, visible failure conditions and a clear distinction between what is supported and what remains unknown.
Evaluation map
Capability, evidence, failure and responsibility must be examined together.
Define what “How should agents divide work, negotiate shared state and verify one another’s results?” includes, what it excludes and which people, tools, organisations or physical conditions form part of the operating system.
Open depth page →Prefer direct experiments, operational records, standards and attributable primary sources. Current reference anchors include NIST, Stanford HAI, Singapore IMDA.
Open depth page →Test not only nominal performance but ambiguity, changing conditions, misuse, dependency failure, human intervention and the route back to a safe state.
Open depth page →Connect findings to the field’s central proposition: From capable models to dependable systems that can act. State what a reader can decide now and what remains premature.
Open depth page →Evidence anchors
These independent sources inform the research frame. Their inclusion does not imply collaboration, review or endorsement.
Interoperability, open protocols, security and agent identity.
Open canonical source ↗02Stanford HAIEvidence that more agents do not automatically produce better outcomes.
Open canonical source ↗03Singapore IMDAPractical boundaries, safeguards and human accountability.
Open canonical source ↗Research protocol
The protocol is designed for maintained research: conclusions can strengthen, narrow or change when new evidence arrives.
State the claim, unit of analysis and operating context before selecting evidence.
Collect attributable primary sources and distinguish direct evidence from adjacent analogy.
Compare supporting and disconfirming results, including negative and null findings.
Test boundary conditions, recovery paths and human or institutional responsibility.
Publish the remaining unknowns, review triggers and reasons a conclusion may change.
FUURAA analysis
FUURAA’s current view is that “How should agents divide work, negotiate shared state and verify one another’s results?” should be judged as a system property, not a headline capability. Progress becomes meaningful when claims can be tied to operating conditions, observable evidence, accountable human decisions and recovery when assumptions fail. The strongest next step is therefore not a larger claim, but a better-defined evaluation that another team can inspect and repeat.
FUURAA Frontier Research™