Anthropic raises agent skills to 500 per session in a major platform upgrade
新智元 · wechat · 2026-07-23
Anthropic shipped six updates to its managed agent platform, pushing the system toward an OS-like agent runtime. The biggest change raises the Skills limit from 20 to 500 per session, enough to encode an entire business unit’s knowledge, policies, and workflows inside one managed sandbox.
Other changes include five effort levels per agent, one-step session creation with up to 50 initial events, thread-level streaming for sub-agents, seven new webhooks covering environment and memory-store lifecycles, and making session version fields optional for simpler single-flow updates. The article frames this as Anthropic betting on deep agent infrastructure rather than consumer app distribution: OpenAI on the entry point, Google on infra, and Anthropic on control depth and operationalized agents.
More from coding & agent
- Pond is hiring a Head of Engineering for an AI agent marketplace in San Francisco — ThePeterMick · 2026-07-23
- Grok Build adds Permissions, Plan Mode, Sandbox and Subagents docs — elonmusk · 2026-07-23
- Cloud agents should run on a persistent machine, not a new sandbox for every PR — aidenybai · 2026-07-23
- ChatGPT/Codex Computer Use is praised for app testing, while a chart puts Kimi K3 at the low-cost end — AxSaucedo · 2026-07-23
- An open-source Agent Skill project claims 100+ templates and 126,000+ GitHub stars — Saboo_Shubham_ · 2026-07-23
- ChatGPT and Codex Computer Use can test apps end to end and use software without APIs — gabrielchua · 2026-07-23