Poolside 2026.03.24

March 24, 2026 Cloud

​

Summary

This release improves how you navigate and manage resources in the Poolside Console, with better visibility into agents and more flexible model configuration. You can use console search to find what you need, monitor inference performance with new agent metrics, export trajectory data for debugging, and track configuration changes over time. It also streamlines model configuration with connection testing and reusable model provider settings, and adds search to the Poolside Web Assistant so you can quickly find conversations and navigate to related resources.

​

New features

​

Poolside Console

• Console homepage and quick navigation
  You can now start from a homepage that gives you quick access to common tasks and recent agent activity. Use Search & Actions to find resources, run actions, and move around the console, Create New to start setup tasks, and Top Agents to see which agents are most active in your environment.
• Custom product branding
  You can customize how Poolside looks for your organization. In Settings > Product Branding, upload your logo, add partnership text, and configure a banner message for the Poolside Web Assistant. Changes apply across your organization.

​

Agents

• Inference metrics for agents
  You can now view inference performance directly in agent dashboards. See time-to-first-token (TTFT) latency metrics, including average, P50, and P95 values, to understand performance and spot potential saturation of inference endpoints.
• Agent trajectory export
  You can export agent trajectory data as a JSON file from the agent detail page. Share this file with Poolside support to help analyze behavior and troubleshoot issues.
• Tool time distribution view
  The trajectory viewer now includes a Time Distribution tab that groups similar tool calls, helping you identify long-running commands, inference-heavy operations, and repeated tool usage during a run.
• Agent configuration history
  You can track changes to agent configurations over time. See what changed, who made the change, and when it happened, and restore previous configurations if needed.
• Try agent dialog
  You can use Try agent to get instructions, commands, and configuration snippets for running an agent across supported surfaces, including the Poolside Agent CLI, IDE extensions, Agent Client Protocol (ACP)-compatible editors, and remote sessions.

​

Models

• Model connection testing
  You can now test a model configuration before saving it. Use Test Connection to verify that the endpoint is reachable and that the model can successfully run inference.

​

Poolside Assistant (IDE)

• More flexible sandbox configuration
  You can configure sandboxes with volume mounts, workspace-specific settings, and a more detailed selector. This makes it easier to reuse local caches and configuration files while keeping projects isolated.
• Sandbox session restoration
  Your sandbox sessions can now be restored when you switch conversations or restart your IDE, so you do not need to rebuild environments for related work.
• Expanded auto-approval controls
  Administrators can allow you to run trusted workflows with fewer prompts by granting the Auto Approve Commands permission. When allowed, you can turn on automatic approvals in the IDE or use --unsafe-auto-allow in the CLI.

​

Poolside Web Assistant

• Search in Poolside Web Assistant
  You can now search to quickly find previous conversations and navigate related resources. Open search from the sidebar or use Command+K (macOS) or Ctrl+K (Windows and Linux). From search, you can reopen conversations, access settings and API keys, or navigate to the Poolside Console.

​

Platform and infrastructure

• Improved model serving
  Model deployments now support improved model reconciliation, authenticated serving connections, and direct stop-token handling.

---

​

Beta features

• Model providers
  You can define reusable connection settings, such as base API URLs and HTTP headers, that multiple models can share. You can also upload custom provider logos, helping you keep model configurations consistent across your organization.
• Knowledge bases
  You can give agents access to read-only knowledge sources backed by Git or Amazon S3. Agents can retrieve and use this information through indexed full-text search.
• Execution environments (sandboxes)
  You can run agents in isolated, disposable container environments that separate execution from the host system while maintaining access to your workspace.
• Agent Client Protocol (ACP)
  You can use agents directly from ACP-compatible editors such as JetBrains, Zed, and Neovim. Connect using the pool acp command to send prompts, receive responses, and run agent workflows from your development environment.
• Kubernetes and OpenShift deployment
  You can deploy Poolside on upstream Kubernetes or Red Hat OpenShift clusters, supporting environments that already use container orchestration.

---

​

Known issues

​

Models

• Instruction-following limitations in Malibu 2.2
  Malibu 2.2 improves instruction following compared to 2.1, but may still struggle with complex or layered custom system prompts compared to other models. Workaround: Refer directly to critical instructions in the user prompt.
• Potential tool-call looping in long sessions
  Malibu 2.2 may enter a looping state where the same tool call is repeated. This has been observed more frequently with long custom instruction sets, large numbers of MCP tools, or extended sessions.
  Mitigation: Guardrails are in place to detect some looping patterns and interrupt execution with a warning.

​

Poolside Console

• Multiple open Poolside Console tabs may cause authorization issues
  Opening the Poolside Console in multiple browser tabs simultaneously may result in authorization errors.
  Workaround: Use a single tab when working in the Poolside Console.

​

Poolside Assistant (IDE)

• Apply Changes may fail after multiple edits in sandboxed sessions (beta)
  In some cases, applying changes from a read-only sandbox may fail after follow-up edits in the same conversation.
  Workaround: Apply changes at the end of a conversation before starting a new one.
• Local sandbox execution cannot be strictly enforced (beta)
  It is not possible to restrict an agent to run exclusively with local sandboxing enabled. Users can still run the agent without sandboxing.
• Binary files may appear in diffs in sandboxed sessions (beta)
  Binary files created in read-only mode may be incorrectly rendered or included in line-count diffs during change review.
  Workaround: Ask the agent to remove built binaries before using Apply Changes, or add them to .gitignore.
• Limited tooling in default sandbox images (beta)
  The bundled default images provide a minimal toolset and may not include commonly expected development tools.
• Private container registries require pre-authentication (beta)
  Poolside Assistant cannot negotiate container registry authentication. Runtimes must already be authenticated to pull from private registries.
• Incorrect approval labeling in sandbox mode (beta)
  Some tool approvals in sandbox mode may be incorrectly labeled as user-enabled unsafe auto-allow.

​

Platform

• Trajectory storage growth
  All execution trajectories are stored in the database, which can lead to rapid database growth over time.
  Workaround: A pruning script is available for administrators to manage storage usage.

​

Component versions