Skip to content
PRESHaiPRESHaiPRESHos Platform

Agent config

Agent config is what the composer Agent picker selects. The profile names the agent, types it for the Agent Ops harness, describes its purpose, and carries its avatar and active status. The runtime policy sets the default model and reasoning plus run bounds such as max steps. Around that core, the surrounding nav attaches the rest: environments, knowledge stores, subagents, tools and plugins, AI providers, MCP servers, skills, memory, models, plugins, schedules, secrets vault, channels, access, and identity.

Config shapes every session that picks the agent, so changes here propagate. Narrow the tool set to what the workload needs, verify before opening up, and keep the FDE probe sequence in mind: reads first, drafts before executes, gates firing where risky work starts.

  • Profile first: name, type, description, and status should read clearly to anyone picking this agent in the composer.
  • Runtime policy bounds the run: default model and reasoning, plus step limits that keep costs and blast radius contained.
  • Capabilities attach deliberately: each tool, MCP server, skill, and knowledge source widens what sessions can reach. Add with evidence, not just in case.
  • Verify after every change with a test session before pilot users arrive.
Default Agent configuration screen in AI Agents

Config checklist

Section titled “Config checklist”
  • Context covers the workload’s systems and nothing adjacent “just in case”.
  • Tool set minimal: reads needed now, drafts for consequential outputs, executes only where the workload requires.
  • Approval gates route to named roles, not named people. See Approvals design.
  • A second person reviewed the config before pilot users entered.
  • Connector-backed tools confirmed against the tenant’s live tool list. Code is not connected.

Scoped helpers owned by this config live under Subagents. Exact creation mechanics (console/API calls, schema) beyond this page are not documented in source yet; treat the product UI as the source of truth for fields this page does not name.

What next

Section titled “What next”