Cohesity Extends Cyber Recovery to AI Agent State

Key Highlights

• Agent Resilience maps agent topology and connects recovery of agent state with the enterprise resources agents depend on and modify.
• Amazon Bedrock AgentCore and Amazon Bedrock Agents are supported first, with Microsoft and Google agent platforms on the roadmap.
• Agent Resilience is available to select customers, with general availability targeted for the end of 2026.
• Gaia Catalog, Agent Resilience, and Autonomous Cyber Resilience connect governed AI data, agent protection, and AI-assisted recovery.

The News

Cohesity introduced Agent Resilience at Catalyst 2026, adding protection and recovery for the infrastructure behind enterprise AI agents. The initial release protects agent memory and configuration, maps agent topology, and connects recovery of agent state with resources agents manage. Initial support is for Amazon Bedrock AgentCore and Amazon Bedrock Agents, with general availability targeted for year-end. For more information, read the official company press release.

Analyst Take

AI agents create a recovery requirement beyond conventional application and data protection. An agent can accumulate memory, hold credentials, invoke tools, and change enterprise systems. In our view, agent state becomes part of the application recovery boundary as agents assume responsibility for business processes. Cohesity defines agent state as more than memory and configuration. Workflows, tool connections, and data stores also influence how an agent behaves. The first release focuses on memory and configuration.

The HyperFRAME Research Lens: State of the Enterprise Infrastructure & Operations (1H 2026) survey found that only 30% of enterprises are very confident they could recover from a major cyberattack or cloud data loss with minimal downtime, and 17% are not confident. Agent state adds another recovery object to an environment where confidence is already limited.

Agent recovery is already emerging as a distinct cyber resilience category. Rubrik reached the market first with Agent Rewind, introduced in 2025 and now part of Rubrik Agent Cloud. Its recovery model traces agent actions, identifies the resulting blast radius, and selectively reverses changes to affected applications and data. Rubrik extended that model in June with configuration backup and restore for Anthropic’s Claude Code agents. Cohesity treats the agent itself as the recovery object, restoring memory and configuration while mapping the resources needed to return it to a trusted state. We believe the two models address different parts of the same recovery requirement.

A single agent may alter one record, several applications, or hundreds of connected resources. Cohesity expects audit trails and higher-level control systems to help identify the appropriate recovery point and scope. Forty-five percent of respondents in the HyperFRAME Lens I&O survey identified lack of visibility and difficulty determining root cause as a very significant outage pain point. Agent actions add more state changes and dependencies that teams must reconstruct before deciding what to restore. In our view, agent recovery will require trusted state plus enough execution history to identify the correct recovery boundary.

Gaia Catalog brings protected enterprise data into the same architecture. It classifies unstructured data, supports SQL and natural-language queries, and lets users create governed datasets for AI and analytics platforms. Cohesity is also taking a selective approach to classification: customers can target specific data sources for deeper, use-case-driven analysis, and broader discovery identifies sensitive data and critical assets at a lighter processing level across the estate. Cohesity describes Gaia Catalog as addressing the upstream AI lifecycle gap, with Agent Resilience addressing recovery after agents begin acting on enterprise data.

Autonomous Cyber Resilience applies agents to the recovery process itself. Cohesity plans to use agentic workflows to assess resilience, recommend policy changes, analyze incidents, and prepare recovery environments. The full autonomous model remains a vision.

Amazon Bedrock is the only agent platform supported initially. Cohesity estimates the market may require support for 20 to 25 agent frameworks, with each integration dependent on APIs that expose enough state for protection and recovery. Connector development will determine how quickly Agent Resilience expands beyond Bedrock.

What Was Announced

Cohesity Agent Resilience discovers the topology around an AI agent and maps the resources it needs to operate. The view shows memory stores and connected systems, along with protection status, so administrators can identify gaps before recovery is required. The first release protects agent memory and configuration using Cohesity’s snapshot architecture, immutable backups, and clean-room recovery. Administrators can select a known-good point in time after memory corruption, configuration errors, or malicious activity, with human approval and object-level selection for protection and restore. Cohesity research found that 56% of organizations are not well prepared to detect, contain, and recover from unintended or incorrect actions taken by AI agents, copilots, or AI workflows.

Recovery can extend into resources changed by the agent. Cohesity already protects many of the databases, file systems, and cloud services agents may modify, allowing agent state and affected enterprise resources to be recovered within the same protection environment. Amazon Bedrock AgentCore and Amazon Bedrock Agents are supported first. Cohesity built an abstraction layer around common agent state so additional platforms can use the same protection framework through platform-specific connectors. Microsoft and Google agent platforms are on the roadmap.

Autonomous Cyber Resilience is Cohesity’s vision for applying agentic workflows to recovery orchestration. Administrators will be able to express recovery objectives and testing requirements through Cohesity Copilot, with Data Cloud assessing posture, recommending policies and recovery rehearsals, and presenting changes for approval. During an incident, the planned workflows can analyze impact, identify attacker activity, and prepare an isolated recovery environment. Cohesity Data Cloud Enterprise Edition and DSPM customers can already automate protection for newly discovered sensitive data, while RecoveryAgent orchestrates portions of incident response and recovery today. Expanded Maestro capabilities expected later in 2026 are designed to connect Cohesity’s protection and recovery functions with Claude, ChatGPT, Gemini, and Helios.

Looking Ahead

Enterprises will need to restore more than the data underneath an agent. They will need the state that determines how the agent behaves, the resources it changed, and enough dependency information to return the business process to a trusted operating condition. Agent Resilience begins to make that state discoverable and recoverable.

Those recovery decisions will depend on evidence from the systems around the agent. Identity records, audit history, transaction data, and agent control systems can help establish what changed and how far recovery should go. Cohesity’s Autonomous Cyber Resilience direction points toward using that context to assess readiness and coordinate recovery, while reserving approval for actions with material business or cost impact.

Gaia Catalog gives Cohesity a path into governed AI data preparation. Connecting that capability with agent protection and recovery through a common control plane could give enterprises continuity from data preparation through AI execution and recovery.

Enterprises evaluating this market should assess more than backup coverage for an individual agent platform. They should examine the depth of agent state captured, the granularity of recovery, dependency awareness, framework coverage, and coordination with enterprise applications and identity. Those capabilities will determine whether agent recovery can support production AI systems at scale.