Formalize or Lose It: What Executable Constraints Control in an Agent Swarm
Multi-agent systems distribute execution across agents, creating more opportunities for requirements to disappear during handoffs. We study how executable constraints can preserve coordination requirements while retaining room for model planning. Drawing on design by contract and runtime verification, we distinguish requirement transmission, plan selection, and host acceptance. We examine how check coverage determines the scope of control through a supervision inventory, fault injection in the Feasibility Scheduler (FBS), and a controlled requirement-relay study. A recorded FBS run that passed all 125 tests carried 84 supervisory rules, of which 39 were communicated only through prompts. Of 23 rules passed from Root plans to Workers, 19 had no executable check. Across 19 planted faults, FBS blocked every plan and delivery fault but accepted six violations outside the effective coverage of its checks. Six additional host checks detected five of those six survivors. In the relay study, a reviewer with full notes corrected every final violation, whereas a reviewer using compacted notes approved 71% of flawed attempts. A ledger delivered requirements directly to the executor and maintained its covered predicates at about one sixth of the supervisory input tokens used by full-note review. Violations involving omitted requirements and unprotected conditional premises still passed acceptance. Our central conclusion is that multi-agent oversight should rest on executable requirements held by the host. Models plan within the permitted space; the host determines which plans and deliveries can be accepted. Reliability depends on whether critical coordination rules reach execution and acceptance. Test pass rates therefore need to be interpreted alongside formalization coverage. This provides a common basis for preserving planning discretion, reducing dependence on compressed context, and designing hybrid oversight.
Authors
- Soumyajit Karmakar (ORCID: https://orcid.org/0000-0003-4867-0970)
- Almene De Meran Meguimtsop
- Samrudh Modalur Radha Krishna
- Nimay Shah
- Sen Lin
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24292368
- Primary Topic
- Multi-Agent Systems and Negotiation
- Type
- preprint