External Assistants¶
Use RunAgents with external assistants (Codex / Claude Code) without relying on built-in RunAgents Copilot.
1. Set External Assistant Mode¶
This keeps RunAgents as the control plane (validation, governance, deployment), while your external assistant handles reasoning.
2. Export Workspace Context¶
This snapshot includes:
- agents
- tools
- model providers
- policies
- identity providers
- approval connectors
- approvals
- deploy drafts
Use --strict if you want export to fail on any partial error:
3. Ask Assistant to Generate an Action Plan¶
Prompt your assistant with context.json and ask it to create plan.json using the Action Plan schema.
Then validate:
4. Apply Plan¶
Add workflow skills when the task is non-trivial¶
For production work, pair exported context and action plans with the public RunAgents skills library. The skills help assistants reason about catalog deployment, approvals, tooling, debugging, and interfaces such as WhatsApp or Slack with much less prompt churn.
See RunAgents Skills.
Recommended Loop¶
context export- assistant generates/updates
plan.json action validateaction applycontext exportagain for confirmation
Safety Guidelines¶
- Always run
validatebeforeapply. - Keep a stable
idempotency_keyper action intent. - Use
continue_on_error: falsefor production deploy sequences. - Store plan files in version control for auditability.