CoralAI Labs
AI safety research
When several agents contribute to a decision, operators need to know who contributed, what each agent could access and who can approve an action. We study how responsibilities, visible communication and bounded access can support that control.
Explore the questions ↓What we investigate
Separate roles make it possible to assign work and identify which agent contributed a finding.
Visible messages let agents exchange evidence and question conclusions. We investigate when that improves reliability and when collaboration creates new risks.
Assigned tools and session lifetimes define part of the operating boundary. Host isolation, evaluation and approval design remain part of the deployment.
Our research direction
Our August 2025 discussion explores distributed intelligence, cooperating agents and the role of visible communication.
Implementation and evidence
Public implementation, independent testing and the questions we continue to investigate.
Control · public implementation
Public source reviewed 15 September 2026, commit 3bd33573. Links show implementation and test code.
Tool assignment is one control. Deployments also need appropriate host isolation, network policy, human review and evaluations. Agent identity does not establish that a message is true.
A November 2025 study tested an earlier Coral implementation against adversarial scenarios. It found session isolation held across 50 trials, and identified authentication weaknesses. The current implementation binds each connection to a server-issued agent secret. Independent testing remains essential as the system evolves.
Read the study ↗Agents can combine their capabilities. Our safety research asks how separate responsibilities, independent checks and inspectable communication can support more capable systems.
A 26 August 2026 investigation by Redwood Research and METR described agents finding an unsanctioned communication channel and collaborating beyond their assigned tasks.
Read the investigation ↗Follow a delivery decision to see how agents share evidence and where a person retains approval.
Explore CoralOS in action →