Anthropic at this time launched Claude Managed Brokers as a analysis preview. Dreaming extends reminiscence by reviewing previous classes to search out patterns and assist brokers self-improve. We’re additionally making outcomes, multiagent orchestration, and webhooks obtainable to builders constructing with Managed Brokers. Collectively, these updates make brokers extra succesful at dealing with complicated duties with minimal steering.
Construct self-improving brokers with dreaming
Dreaming is a scheduled course of that critiques your agent classes and reminiscence shops, extracts patterns, and curates recollections so your brokers enhance over time. You determine how a lot management you need: dreaming can replace reminiscence mechanically, or you may evaluation modifications earlier than they land.
Dreaming surfaces patterns {that a} single agent can’t see by itself, together with recurring errors, workflows that brokers converge on, and preferences shared throughout a group. It additionally restructures reminiscence so it stays high-signal because it evolves. That is particularly helpful for long-running work and multiagent orchestration.
Collectively, reminiscence and dreaming kind a strong reminiscence system for self-improving brokers. Reminiscence lets every agent seize what it learns as it really works. Dreaming refines that reminiscence between classes, pulling shared learnings throughout brokers and protecting it up-to-date.
Dreaming is out there in Managed Brokers on the Claude Platform; builders can request entry right here.
Ship higher outcomes
With outcomes, you write a rubric describing what success seems to be like and the agent works towards it. A separate grader evaluates the output towards your standards in its personal context window, so it isn’t influenced by the agent’s reasoning. When one thing isn’t proper, the grader pinpoints what wants to alter and the agent takes one other cross.
Brokers do their finest work after they know what “good” seems to be like. For instance, a structural framework, a presentation customary, or a set of necessities that should be met. With outcomes, brokers can examine their work towards that bar and self-correct till the output is nice sufficient, with out a human needing to evaluation every try.
Outcomes is especially helpful for duties that require consideration to element and exhaustive protection. It additionally works for subjective high quality, like whether or not copy matches a model voice or a design follows visible pointers. In testing, outcomes improved process success by as much as 10 factors over an ordinary prompting loop, with the most important positive aspects on the toughest issues. Outcomes additionally improved file era high quality, with +8.4% process success on docx and +10.1% on pptx in our inside benchmarks.
You can too now outline an end result, let the agent run, and get notified by a webhook when it’s finished.
Deal with complicated duties with a number of brokers
When there’s an excessive amount of work for a single agent to do properly, multiagent orchestration lets a lead agent break the job into items and delegate every one to a specialist with its personal mannequin, immediate, and instruments. For instance, a lead agent can run an investigation whereas subagents fan out via deploy historical past, error logs, metrics, and assist tickets.
These specialists work in parallel on a shared filesystem and contribute to the lead agent’s general context. The lead agent can examine again in with different brokers mid-workflow as a result of occasions are persistent and each agent remembers what it’s finished. You can too hint each step within the Claude Console: which agent did what, in what order, and why, supplying you with full visibility into how your process was delegated and executed.