a day ago
- Managed agents are cloud-hosted agent harnesses (e.g., Claude Code) that run 24/7, allow background operations, and provide security and sandboxing benefits.
- Vendor lock-in is a major risk with managed agents; switching between agent harnesses (e.g., Claude Code to Codex) is easy locally but becomes difficult once deeply integrated into a provider's cloud.
- Anthropic's recent pricing changes for non-interactive Claude Code usage (5-20x increase) highlight the danger of lock-in and give OpenAI an opening with Codex's more flexible allowances.
- Solutions to avoid lock-in include self-hosting managed infrastructure (e.g., running Docker with OpenCode) or using third-party platforms that support multiple model providers.
- The author recommends self-hosting for now to build organizational competence, but warns that frontier labs may introduce exclusive capabilities only available on their managed agent platforms.