I’m running into an issue with the Agentic Runtime for Claude Managed Agents / OpenClaw, and wanted to flag it.
According to the documentation,
- Web browsing — agents can search and browse the web in real time
- PDF creation — generate documents on the fly during a conversation
- Code execution — run code directly from the agent’s sandboxed computer
- Scheduled & recurring tasks (cron jobs) — agents can run on a schedule, kicking off tasks automatically at set intervals (hourly, daily, weekly, or any cron expression) without anyone needing to start the conversation
- Computer access — request a link to view and control the agent’s computer, allowing you or your users to authenticate accounts, verify actions, or interact with the environment directly
What I’m actually seeing is quite different. When I prompt the Agentic Runtime to use an installed action (e.g. Gmail or a Composio action), it doesn’t send me an OAuth login link the way I’d expect. Instead, it just responds that it “cannot perform this action” — with no prompt to authenticate, and no error explaining why.
By comparison, the non-Agentic Runtime handles this correctly and behaves as expected. Right now the Agentic Runtime feels significantly less capable in practice — more like a basic chatbot than an agent that can actually execute actions.
Steps to reproduce:
- Install a Gmail or Composio action on the Agentic Runtime.
- Prompt it to perform that action.
- Instead of receiving an OAuth/login link, the response is “It cannot perform this action.”
Could you help me understand:
- Whether this is expected behavior for the Agentic Runtime, or a bug
- Why OAuth links aren’t being surfaced for Gmail/Composio actions specifically
- Whether there’s a config step I’m missing (e.g. entity ID, connector scope) that would fix this
Happy to share screenshots, logs, or example prompts/responses if that helps debug.