Anthropic has released a mid-cycle update to Claude Sonnet 4 that adds two major capabilities that enterprise customers have been waiting for: production-stable computer use and a formal multi-agent coordination interface that removes the need for custom orchestration frameworks.
Computer use — the ability for Claude to directly control a desktop or browser environment by taking screenshots, clicking, typing, and running terminal commands — was introduced in beta in late 2025. The July 2026 update graduates it to production status, with Anthropic publishing a formal SLA and a suite of safety constraints that are configurable at the API level. Enterprise customers can now specify a "permitted action surface" — limiting Claude to operating within specific applications or URL patterns — reducing the risk of out-of-scope actions during autonomous task execution.
The multi-agent coordination feature is more architecturally significant. Previously, developers who wanted Claude to spawn sub-agents for parallel workstreams had to build custom orchestration code. The updated API now includes a native "delegate" tool that allows a supervisor Claude instance to describe a subtask, hand it to a fresh Claude context with a specific instruction set, receive the result, and incorporate it into its own reasoning — all within a single API session at the orchestrating layer.
In internal benchmarks, multi-agent Claude architectures using the native delegation tool completed complex research-and-draft tasks 3.2x faster than single-agent approaches, with higher factual accuracy because specialised sub-agents can be given narrower instruction sets that reduce hallucination surface.
Anthropic's head of product noted that the delegation protocol includes cryptographic audit logging so enterprises can trace exactly which sub-agent produced which portion of a final output — a critical requirement for regulated industries where AI decisions must be explainable and attributable.
The update also improves Claude's performance on long-document tasks, with retrieval accuracy on 100k-token documents improving from 78% to 91% on Anthropic's internal NIAH (Needle in a Haystack) evaluation. For legal and financial services customers processing lengthy contracts or filings, this closes a gap that previously made 100k+ context windows unreliable in practice.
Pricing for Sonnet 4 remains unchanged at $3 per million input tokens and $15 per million output tokens, with computer-use and multi-agent sessions consuming tokens from both the supervisor and sub-agent contexts. Anthropic provides tooling to estimate session costs before execution.