The ClaudeDevs account outlined a batch of additions to Claude Managed Agents on Wednesday, including "multiagent orchestration," an "outcomes loop" for rubric-driven self-improvement, "dreaming" for self-learning, and webhook support. The announcement post had drawn 4,722 likes, 343 reposts and 123 replies.
In the thread, ClaudeDevs stated that multi-agent orchestration lets tasks be delegated to specialized sub-agents. Each sub-agent keeps its own context window while sharing a container and filesystem, which the account describes as a way to coordinate work across agents.
The more consequential additions appear to be the feedback and memory systems. ClaudeDevs says agents can self-improve using a rubric that defines task completion, while a dedicated grader sub-agent evaluates the work and returns feedback for the next iteration. Another post states that agents can publish updates to webhooks configured in a Claude workspace, removing the need to keep an SSE stream open or poll for completion. The account also describes "dreaming" as a background process that reviews many sessions to curate memory, including editing memories based on patterns, adding new skills or removing stale ones.
ClaudeDevs also pointed users to the /claude-api skill in Claude Code or the OSS repo to get started. Replies on X quickly turned to cost, limits and control, with commenters asking whether the agents would be cheaper, whether subscriptions would cover them, and how the system would avoid optimizing for a rubric while drifting away from the actual goal.
Source: ClaudeDevs on X
