Anthropic supercharges Claude Code and rolls out multi-agent tools

If you use AI to write code, you know the two biggest annoyances: having to re-teach it your repo's quirks every morning, and watching it drift off into hallucination land during long, multi-file edits.

At its Code w/ Claude developer conference in San Francisco, Anthropic rolled out a slate of updates to Claude Managed Agents that tackle both problems directly.

Agents that sleep on it and work in packs

The most intriguing new feature is called Dreaming.

Just like the human brain consolidates memories while sleeping, this background process combs through previous agent sessions to find patterns. It catches recurring mistakes, documents workflows, and curates long-term memory so your agent actually learns team preferences between runs.

Anthropic also added native multiagent orchestration.

A single lead agent can now spin up specialized subagents that work in parallel across a shared filesystem. Each subagent can use its own model, prompt, and toolset. The entire execution trace is visible inside the Claude Console, and new Webhooks notify you the second the job wraps up.

Bringing in a strict grader

To make sure all this parallel code does not turn into a mess, Anthropic introduced Outcomes.

Instead of trusting an agent to grade its own homework, developers can now define a strict rubric. A dedicated grader evaluates the work in an isolated context window and kicks it back to the agent for revisions until it passes. Anthropic claims this closed-loop feedback lifted task success rates on its toughest internal benchmarks by up to 10 points.

On top of that, Anthropic doubled rate limits for Claude Code and raised API limits for Claude Opus, making heavy-duty automated development much more viable.

Why it matters

Anthropic is tackling the actual plumbing that keeps agentic software from shipping to production.

For developers and engineering leads, this is worth testing immediately. Building parallel agents, memory compaction, and external evaluators from scratch usually requires hundreds of lines of fragile glue code. Having Anthropic handle that infrastructure out of the box cuts massive friction.

It also shifts the competitive dynamics. Companies like Cursor, GitHub, and Replit are Anthropic customers, but they are also building proprietary agent flows. By baking orchestration, automated curation, and graders directly into the Claude Platform, Anthropic is turning foundational infrastructure into a managed service.

The catch is cost. Running subagents in parallel with an independent grader constantly rejecting work will burn through tokens fast. If you don't keep a close eye on execution loops, your monthly API invoice might come with serious sticker shock.