Research programmes
Experimental environment
A bounded cooperative of agent teams.
A research proposal comparing institutional coordination, decentralised protocols and task orchestration when AI agent teams share resources.
Research proposal
This programme proposes a bounded environment in which teams share compute slots, data-access budgets and task queues. The comparison asks whether explicit commitments and review improve coordination enough to justify their cost.
This brief sets out a question and proposed method. It does not report experimental findings. Etylon project · Updated 20 October 2026.
Explicit commitments, bounded authority and reviewable allocation should improve long-horizon coordination when resource contention and conflicting goals become material.
Research setting
One environment. Several teams. Contested resources.
The boundary makes coordination gains, governance cost and recovery from conflict observable rather than anecdotal.
Proposed research protocol
Intervene narrowly. Compare longitudinally.
- 01Baseline
Compare message-only teams, decentralised protocols, orchestration and simple allocation rules.
- 02Shadow
Run institutional checks without changing decisions.
- 03Calibrate
Analyse false holds, misses and governance labour.
- 04Intervene
Gate a narrow class of reversible actions.
- 05Compare
Report benefits, costs and failure distributions.
A case to examine
An urgent task competes with a reserved compute run whose output another team needs. The institutional condition would expose the existing commitment and support a proposed capacity swap. The comparison would examine completed work and disrupted dependencies, including cases where the urgent task should wait.
What to compare
Compare institutional coordination with message-only teams, explicit decentralised protocols, a competent task orchestrator and simple allocation rules. Hold models, tools, information access and resource budgets comparable. Vary goal conflict, delayed information and changing participation; report outcomes, overhead and who bears the costs.
Conditions and limitations
Specify the unit of comparison, budgets and success criteria before evaluating outcomes. Report denominators, variability and adverse cases alongside averages. Shadow checks reveal proposed holds but do not establish the effects of enforcement; those require a separate intervention. Benefits in a bounded resource setting would not establish benefits for every multi-agent system.
Proposed measures
Define the evidence before judging the result.
- Coordination gain
Change in completed shared work and missed commitments relative to the specified baseline.
- Duplicate work
Repeated work that provides no additional task value, reported against total work attempted.
- Conflict recovery
Time and additional work needed to resolve incompatible commitments.
- Goal retention
Whether the agreed task objective survives interruptions and repeated hand-offs.
- Review labour
Time spent understanding, reviewing and correcting institutional decisions.
- False holds and misses
Valid proposals held and conflicting proposals passed, judged against the task’s declared criteria.
What would count against the proposition
Institutional procedure adds material latency or review labour without producing a corresponding improvement in coordination, recovery or resource use.