Robert Ta and Jonathan demonstrate Claude's new managed agents, which use rubric-based self-improvement to automatically evaluate and refine work. They stress that high-quality rubrics, built with domain experts like accounting teams, are essential for autonomous agents to deliver real business value.
New Claude Code features in Claude Managed Agents. They've added multi-agent orchestration and outcomes loop for rubric-driven self-improvement and dreaming for self-learning, as well as webhooks. Here's what happened and here's what it means for AI builders. So just reading this real quick, the Claude Devs Twitter handle— I would suggest people follow this— multi-agent orchestration enables task delegation to specialized subagents. Agents can self-improve using a rubric that describes task completion. A dedicated grader subagent then uses the rubric to evaluate any work done, returns feedback for the next iteration, and then agents can post updates to webhooks configured in your Claude workspace, and that removes the need to keep an SSE stream open or poll for completion. Agents write to memory in a session, and dreaming is a background process that reflects over many sessions to curate memories. It can edit them based on patterns, add new skills, or remove stale ones. Jonathan, you mentioned that outcomes are a very interesting thing. Could you say more on that?
We've been talking about agents and agents getting to closed loops. The idea that you can just specify an outcome and then use a subagent as an LLM judge to make sure that the agent is, is producing outputs that are in line towards this outcome is really powerful. I think it's not magic, right? Because you still have to define a really good rubric., and you need to do that outside of what the tools that Anthropic gives you here. You know, we've been taught, and like, this is how we work to, to build datasets, right? To use, uh, human expertise, open coding, to define the axial coding and then the rubric of different aspects that, uh, a, a given agent should conform to. So you can do that outside of the system, but then once you have a good rubric, you can just ask your agent to create the outcome and use LLM as a judge to do that. And so it takes away a lot of the machine learning work and a lot of the engineering work., and it makes it simpler to create agent-based outcomes. So I think a lot of Cloud Code users might look at this as like, um, I'll be able to use this on all of my agents. And it's not like a beginner's tool in my mind. It's, it's more like, um, once you know how to create good rubrics, then you've got a really powerful way to use agents with LLM judges to create outcomes. And I think that's something that's very empowering. Yeah.
I think the implication for AI builders, uh, especially ones building for businesses, right? Um, if, if you're in an enterprise or you're building for an enterprise client as an AI builder, I totally see this as part of the bigger strategy shift around OpenAI and Anthropic competing for the enterprise market. They're trying to ship features that make it easier for the AI common, you know, citizen AI developer or builder to build closed-loop agents. Me and Jonathan talk about this a lot for our own company strategy. Like, what are the moats in the future? It's clear to us that the world is going to move to autonomous closed-loop agents around economically valuable workflows, full stop. And in this, like, I could totally see just them creating this makes that AI builder's life easier to go install a closed-loop autonomous agent around some economically valuable workflow at an enterprise company, right? Because how— the question kind of down to the details becomes, well, if the rubric is the gold, where you need to get a good rubric for a particular domain for that workflow, say here they're using the example of, uh, revenue projections and some financial model, uh, case. That's something every company has to do is close their books and have forecasts and reporting, right? Every company's got to do that, but every company's revenue recognition and their accounting will look different depending on the maturity of the company, depending on the industry, depending on how they want to position for their investment or an IPO or whatever, right? So obviously the rubric is going to be a bit different for a financial agent closing books at Company A versus Company B. So how does a developer who doesn't understand accounting do that? Tactically, what they would do is go talk to the CFO and their accounting team and go translate all the open coding, all the datasets, you know, look at traces, et cetera, get into a failure taxonomy, get to a rubric. What Anthropic did with this move is just make that particular process a lot easier on that developer or builder, uh, which I think is a great thing for their enterprise adoption story.
