Choose an Agent runtime

Overview

The Runtime controls how the Agent executes its instructions and connects to external systems. When an Agent starts, Agent Barn:

1. Loads the Agent and its pinned Template version.
2. Renders the Template files with the Agent's identity.
3. Loads the Agent's pinned Skills.
4. Decrypts the required platform and integration credentials.
5. Builds configuration for Hermes or OpenClaw.
6. Creates or refreshes the Agent's Runtime resources.
7. Connects the Runtime to the selected communication platform.
8. Reports Conversations, Tool calls, Logs, and health back to Agent Barn.

Runtime configuration is regenerated every time the Agent starts.

The Runtime does not define the Agent's purpose. The Template defines its identity and behavior, and Skills add focused instructions and integrations. Template, model, Runtime, platform, and Skills remain separate choices.

Runtime comparison

Neither Runtime is inherently correct for every Agent. Let platform compatibility and the operating controls you need decide the selection.

Platform compatibility

Current web-wizard availability

The current Hire agent wizard offers Slack, Telegram, and Discord.

Microsoft Teams is supported by both Runtimes and by the Agent API contract, but it is not currently offered as a platform choice in the web hire wizard. Provision a Teams Agent through the Agent API, and confirm that the deployment has a supported Teams provisioning path before committing to it.

Choose Hermes

Hermes is the recommended Agent Barn path, and the Runtime the hire wizard preselects.

Choose this when

• This is your first Agent
• The Agent will use Slack, Microsoft Teams, Telegram, or Discord
• You want explicit command approval settings
• You want Slack responses to be Verbose or Concise
• You need a configured home channel for scheduled or proactive Slack or Telegram delivery
• You want the documented Hermes platform-recovery behavior
• You want scheduled runs to load persistent memory without inheriting an interactive chat session

Hermes command approval

Agent Barn exposes three command approval modes for Hermes:

Use Manual for Agents working with sensitive systems or consequential tools. Use Auto for controlled workflows where low-risk commands should proceed without interrupting the user. Use Off only when the surrounding security controls make approval prompts unnecessary.

Hermes Slack verbosity

For a Hermes Slack Agent, choose:

Verbose mode helps during evaluation and troubleshooting. Concise mode is usually more appropriate once the workflow is stable.

Slack Verbosity is not shown for Telegram, Discord, or OpenClaw.

Hermes scheduled work

Hermes scheduled runs use isolated sessions. They do not inherit the Slack thread or interactive session that originally created the schedule, but they do load the Agent's persistent memory and user-profile context.

For scheduled or proactive Slack and Telegram delivery, configure a home or alert destination channel. Without a destination, the Runtime can complete work with no appropriate place to deliver its result.

When a scheduled Hermes run has nothing to deliver, its final response must be exactly:

[SILENT]

Ordinary prose such as "Nothing to report" is treated as a deliverable message.

Hermes platform recovery

Hermes retries eligible platform connection failures with bounded exponential backoff.

If repeated failures open its platform circuit breaker, the Runtime health probe becomes unhealthy and Kubernetes can restart the Agent pod. That gives recoverable platform configuration changes another chance to connect.

An intentionally paused Hermes platform remains live and is not restarted solely because it is paused.

Consider another Runtime when

• The workflow depends on OpenClaw-specific behavior
• Your existing Templates and operational testing target OpenClaw
• Command approval must remain managed by OpenClaw

Choose OpenClaw

Choose OpenClaw when its platform support or Runtime behavior is a specific requirement.

Choose this when

• Your deployment already operates OpenClaw Agents
• Your Template was written and tested for OpenClaw behavior
• You want OpenClaw's broader channel configuration architecture
• You accept command approval being managed by OpenClaw

OpenClaw supports Slack, Microsoft Teams, Telegram, and Discord. The selected platform still requires its own credentials and routing configuration, whichever Runtime runs the Agent.

OpenClaw channel behavior

Agent Barn generates an OpenClaw configuration overlay for the selected platform. The overlay covers:

• Model routing
• Channel binding
• Group and direct-message policy
• Allowed platform identities or locations
• Mention gating
• Memory
• Tool profile
• Telemetry plugin
• Runtime health configuration

For Slack, Agent Barn explicitly configures Socket Mode, channel and direct-message policies, explicit mention requirements, fresh mentions in threads, and streaming behavior. For Telegram and Discord, it generates the corresponding group, user, and channel restrictions. For Microsoft Teams, it generates the corresponding webhook channel integration.

OpenClaw command approval

Agent Barn displays OpenClaw command approval as:

Managed by OpenClaw

The Auto, Manual, and Off selector is not shown for OpenClaw Agents. Choose Hermes if selecting command approval behavior through Agent Barn is an operational requirement.

Consider another Runtime when

• The Agent has no OpenClaw-specific requirement
• You want Agent Barn's user-selectable command approval
• You want Slack Verbose or Concise behavior
• You want the documented Hermes scheduled-session and recovery behavior
• You are following the recommended first-Agent path

What both runtimes provide

Hermes and OpenClaw share the same Agent Barn domain model. Both receive the same inputs and report the same signals.

• The exact pinned Template version
• The Agent's rendered identity and operating instructions
• The configured model and per-Agent LiteLLM configuration
• Exact pinned Skill versions
• Provider integration context
• Platform credentials and encrypted Agent secrets at startup
• Per-start ingest credentials
• Runtime behavior and mention policy

Both create or use managed Runtime resources for configuration, secrets, persistent storage, a network service, the Runtime deployment, and health reporting. Both report inbound and outbound Conversation messages, Tool-call state and results, Runtime Logs, and health status.

Secret values remain encrypted at rest and are omitted from Agent reads under either Runtime.

Mention gating

Both Runtimes require explicit mentions in shared channels and groups. On Slack, both require a fresh mention on every message, including replies inside an existing thread. Direct messages are exempt from shared-channel mention gating when direct messages are enabled.

Runtime selection must not be used to bypass communication boundaries.

Templates and Skills

A Template does not become compatible with every platform merely because both Runtimes can start it. Review the Template's identity, user instructions, tool references, platform language, required Skills, and required providers.

A Template that says it is embedded in Slack, or that references Slack-only tools, remains Slack-oriented under OpenClaw. A Slack-only Skill also remains Slack-only under either Runtime, and platform-specific credentials are required regardless of Runtime.

Operational differences

These implementation differences should not leak into the Template unless the Agent's task genuinely depends on Runtime-specific behavior.

Changing Runtime

Runtime is a creation-time property. The Agent configuration page displays the Runtime as an operational fact, but Agent update provides no Runtime field. Platform is also not changed through Agent update.

To move an Agent to another Runtime or platform:

1. Record the existing Agent's Template key and exact version.
2. Record its model and assigned Skill versions.
3. Record its platform access policy.
4. Identify which credentials can safely be reused.
5. Hire a replacement Agent with the required Runtime and platform.
6. Assign the corresponding Template and Skills.
7. Verify the replacement Agent in a controlled location.
8. Move platform traffic to the replacement.
9. Stop the original Agent.
10. Delete the original only after its history and cost requirements are understood.

Changing a Template, model, Skill, or command approval mode does not change the Runtime, and changing the Runtime is not the same as changing the Template.

Verify the Runtime

After hiring the Agent, confirm:

• The Agent detail page displays the intended Runtime badge
• The selected platform is compatible with that Runtime
• The Agent reaches Working
• An allowed platform mention receives a response
• An unmentioned shared message receives no response
• Conversations and Tool calls appear under the correct Agent
• Live Logs show the expected Runtime starting
• No compatibility or platform-credential error is reported

For the complete procedure, see Verify your Agent.

Troubleshooting

{ "agent_type": "hermes" }

{ "agent_type": "openclaw" }

Next steps

After choosing a Runtime:

• Hire your first Agent
• Verify your Agent
• Configure an Agent
• Manage the Agent lifecycle
• Configure channel access
• Review Agent health and logs