Skip to content

Agent Configuration

Overview

The agent detail page (/agents/:id) provides a complete set of tools to configure model, behavior, sandbox, rate limits, and context files. This page is intended for admins and operators who need to fine-tune agents for production environments.

Route: /agents/:id Sidebar Group: Core Access: Logged in (some sections require Admin)


Agent Detail Interface

The page consists of 4 main tabs:

TabContent
GeneralModel configuration, provider, behavior settings
FilesContext files (SOUL.md, IDENTITY.md, AGENTS.md, ...)
PermissionsPer-user permissions
InstancesOnly visible for predefined agents — list of per-user instances

Main actions on the page:

  • Update agent configuration
  • Regenerate from prompt — rebuild the agent from a new description
  • Resummon — reinitialize the agent (re-run context file generation)
  • Delete agent — confirmation dialog
  • Advanced settings — open the advanced dialog
  • Configure Heartbeat — set up health checks
  • Manage Codex Pool — navigate to /agents/:id/codex-pool

Configuration Guide

Model and Provider

In the General tab:

FieldDescription
ProviderRegistered provider name (e.g., anthropic, openrouter)
ModelSpecific model name (e.g., claude-sonnet-4-5-20250929, gpt-4o)
Max IterationsMaximum loop iterations per run (default: 20). Increase if the agent has many complex tool calls.
History LimitNumber of user turns retained in history. Shorter saves tokens, longer helps retain context.

Context Files (Files Tab)

Files that define the agent's personality and knowledge:

FileFunction
IDENTITY.mdAgent's name, role, background. Injected into the system prompt at the "primacy zone" level.
SOUL.mdMain system prompt — communication style, values, handling principles.
BOOTSTRAP.mdIf present, the agent runs immediately on startup (mandatory notice).
AGENTS.mdDescribes agents in the system so the main agent knows how to spawn sub-agents.
TOOLS.mdDocumentation about tools available to the agent.

You can add any file to the workspace. Files are injected into the system prompt under the "Project Context" section with a defensive preamble that protects the agent from being manipulated by external content.

Per-user context files (open agents): Each user has a separate base/{userID}/ directory. The USER.md file is created automatically when a user chats for the first time — the agent can update this file over time.

Regenerate from prompt (Files Tab):

  1. Click Regenerate from prompt.
  2. Enter the new role description in the textarea.
  3. Click Regenerate — the system regenerates SOUL.md and other context files.

Behavior Settings

SettingDescription
DebounceWait time (ms) before processing messages in group chat (prevents spam)
StreamingEnable/disable streaming response by chunk. Disable for debugging.
Tool StatusDisplay tool call status in the UI (tool.call / tool.result events)
Input GuardActivate prompt injection scanning (default: warn — logs but does not block)

Advanced Settings

Opened via the Advanced button:

SectionContent
Workspace SharingConfigure workspace sharing between agents
ReasoningMode, level, fallback for Extended Thinking
ChatGPT OAuth RoutingCodex Pool routing configuration
Context CompressionThreshold and strategy for conversation history compression
Context PruningRemove less important context sections
SandboxDocker sandbox configuration for code execution

Sandbox Mode

Run code in a Docker container to isolate it from the main host.

FieldDescription
Modeoff — no sandbox; non-main — sandbox sub-agents only; all — sandbox everything
Workspace Accessnone full isolation; ro read-only mount; rw full read-write
ImageDocker image to use (e.g., goclaw-sandbox:bookworm-slim)
Scopesession — 1 container per session; agent — shared across sessions; shared — shared across all agents
TimeoutMaximum time (seconds) for each command run in the sandbox (default: 300)
MemoryMaximum RAM (MB) for the container (default: 512)
CPUsNumber of CPU cores (fractions allowed, e.g., 0.5)
NetworkEnable/disable network access from the container

Sandbox requires Docker installed on the host. The image must be pre-built.

Rate Limits and Permissions (Permissions Tab)

  • Limit the number of tool calls within a time period (per hour/day/week).
  • Apply RBAC: admin / operator / viewer.
  • Deny specific tools for the agent via Tool Policy.

Adding per-user permissions:

  • Enter User ID, Config Type, Scope, Permission (allow/deny).
  • Click Add (+) to grant permission.
  • Click Delete (X) to revoke permission.

Heartbeat (Health Check)

Click Configure Heartbeat to set up:

FieldDescription
Enable/DisableTurn heartbeat on or off
Interval (minutes)Time between health checks
Provider/Model overrideUse a different provider/model for heartbeat
ChannelChannel to receive notifications
Chat IDChat ID to receive notifications
Active hoursTime window for running heartbeat
TimezoneTimezone to apply
ChecklistCustom check steps

Click Run Now to test immediately. View heartbeat history via Heartbeat Logs.


Example — Configuring Sandbox for a Code-Writing Agent

/agents/my-coder -> Advanced -> Sandbox:
  Mode: non-main
  Workspace Access: rw
  Image: goclaw-sandbox:bookworm-slim
  Scope: session
  Timeout: 120
  Memory: 1024
  Network: enabled
-> Save

Notes

  • Changing Model/Provider takes effect immediately for new sessions; running sessions are not interrupted.
  • Setting Max Iterations too low may cause the agent to stop mid-task when there are many complex tool calls.
  • Sandbox requires the Docker daemon to be running on the host — check docker info before enabling.
  • The Instances tab is only visible for predefined agents — it allows viewing and managing each user's individual context.

See Also