Anthropic AI Updates: July 22, 2026

1. Anthropic Details How It Secures an AI-Native Software Development Lifecycle

Anthropic. Deputy CISO Jason Clinton described how the company secures a development pipeline in which Claude now authors about 80 percent of merged code while human engineers approve between 19 and 54 percent of pull requests. The approach shifts security left by encoding guidelines in CLAUDE.md files, contains blast radius with hard identity boundaries, and replaces single super-agents with multiple specialized review agents plus human sampling across the plan, code, test, deploy, and monitor stages. Clinton frames the change as the security engineer moving from monitoring bugs to monitoring loops, and treating agents as potential insider threats with minimal permissions and logged agent-to-agent communication. Source

2. Datadog Has Claude Code Agents Write Verified Specifications Instead of Code

Anthropic. In a Claude Code case study, Datadog described Temper, a deterministic kernel it calls a universal machine tool that has agents generate formal specifications of system behavior, data contracts, and authorization policies rather than writing application code directly. The kernel verifies each specification through four independent layers, symbolic reasoning over guards and invariants, exhaustive state exploration, deterministic simulation with fault injection, and randomized property testing, then runs the verified specification in production to eliminate drift. Datadog uses the setup to run an autonomous dark factory for its Helix streaming service, targeting the gap where agents generate code faster than humans can verify it. Source

3. Claude Code v2.1.217 Adds a Concurrent Subagent Cap and Transcript Write Warnings

Anthropic. Anthropic shipped Claude Code v2.1.217 with several additions aimed at reliability. A new cap limits concurrently running subagents to 20 by default, overridable with the CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS variable, so a single message cannot fan out an unbounded number of background agents. The release also warns when transcript writes fail, for example on a full disk, or when session saving is silently off due to an inherited environment variable, and it adds emoji shortcode autocomplete in the prompt input. Source

4. Claude Cowork Learns New Skills From Screen Recordings

Anthropic. Anthropic’s Claude Cowork desktop app can now capture screen recordings with voice-over narration and convert those demonstrations into reusable workflows the agent can replay on its own. Teaching an agent by showing it a task, rather than scripting each step, lowers the barrier to automating repetitive desktop work for non-developers. The feature points toward demonstration-based onboarding as an alternative to written instructions or code for configuring agents. Source