Anthropic shipped a major update to Claude Managed Agents on July 22, 2026, raising the number of skills an agent can mount in a single session from 20 to 500 and adding a five-tier effort control for its models. For anyone building agents on Claude, this is the difference between a demo and a production system that can hold an entire knowledge base at once.

Try It: Mount More Skills Without Blowing Your Context

The headline change is the skill cap, but the mechanism that makes 500 skills usable is progressive disclosure. Instead of loading every skill's full definition upfront, Claude Skills now use a three-tier, on-demand loading system that compresses the context footprint of 50 skills from roughly 150,000 tokens down to about 2,000. To use it today, pull the latest Managed Agents SDK, attach your skill set to an agent, and set an effort level inside the model object when you create the agent. Start low for routing and classification, and reserve the top tier for genuine multi-step reasoning.

Why It Matters for Builders

Skill sprawl was the practical ceiling on Claude agents. A 20-skill limit forced builders to split workflows across multiple agents or constantly swap skill sets, both of which add latency and failure points. Lifting the cap to 500, and making it affordable through progressive disclosure, means a single agent can now represent a whole department's tools and playbooks. The five-tier effort control is the cost lever that makes this sustainable: cheap routing for simple tasks, deep reasoning only when the job demands it, decided per request rather than per agent.

Key Details

Skills per session: raised from 20 to 500

Progressive disclosure: three-tier loading, 50 skills compressed from ~150K to ~2K tokens

Effort controls: five-tier thinking effort, set in the model object at agent creation

Also shipped: one-step seed session creation (up to 50 events, zero cold start), sub-agent real-time streaming at thread level, and 7 new lifecycle webhooks

Event date: July 22, 2026

What to Do Next

Audit any agent you built against the old 20-skill ceiling and consolidate the workflows you had to split apart. Check the Claude release notes for the exact SDK version, then tune effort levels per task before you scale. If you are still authoring skills, our walkthrough of recording a Claude skill by screen capture is the fastest way to build the library you will now be able to mount in full.