Research
Secure Long-Running Agent Swarms
June 6, 2026 · RNDLabs Oy
Long-running autonomous agents need more than orchestration. They need evidence trails, permission boundaries, isolated execution environments, and human-readable state.
RNDLabs is exploring an Agent-Swarm stack with cryptographically secured audit trails and MicroVM-secured sandboxes for safer task execution.
The problem with long-running autonomy
Short agent demos are easy to inspect. Long-running work is different. An agent may research, modify files, call tools, generate code, analyze data, and coordinate with other agents over many hours. If the system cannot explain what happened, operators cannot safely approve the outcome.
The core design challenge is therefore not just planning. It is evidence. Every important step needs durable context: intent, inputs, tool calls, outputs, model decisions, policy checks, and human approvals.
Trust boundaries
A secure agent-swarm stack should separate several responsibilities:
- The operator defines task intent, approval requirements, and allowed resources.
- A planner decomposes work into bounded steps.
- Specialist agents handle research, coding, security review, data analysis, and operations.
- Tool execution happens inside constrained environments, ideally MicroVM-backed for risky code paths.
- Audit infrastructure records what happened in a way that can be inspected, replayed, and challenged.
Auditability as a product feature
Audit logs should not be an afterthought. For industrial and B2B use, auditability is part of the product. It supports governance, debugging, customer trust, incident response, regulatory evidence, and model-quality improvement.
RNDLabs is especially interested in cryptographically secured audit trails: logs where important events can be chained, signed, or otherwise made resistant to silent modification. The exact implementation depends on deployment constraints, but the operating principle is stable: autonomous work should leave durable evidence.
Research direction
Current research questions include sandbox escape resistance, model/tool policy enforcement, human-in-the-loop review design, model audit workflows, evidence compression, secure retrieval, and how to keep long-running agent state understandable to a human operator.