The most interesting development in AI tooling this week isn't a new model — it's the supporting infrastructure maturing around agents. Anthropic shipped persistent agent memory and overhauled its Managed Agents platform, OpenAI published a Claude Code plugin that routes tasks to Codex, and the two ecosystems are quietly becoming more interoperable. The stack required to trust agents with real, long-running work is clicking into place.
OpenAI Publishes a Codex Plugin for Claude Code
OpenAI released codex-plugin-cc, a Claude Code plugin that brings Codex's cloud inference into Claude Code sessions without requiring a context switch. Three commands cover the primary use cases: /codex:review reviews uncommitted changes or branches, /codex:adversarial-review goes further and challenges design decisions and implementation choices rather than just flagging bugs, and /codex:rescue (or /codex:transfer) hands a task off to a background Codex job for investigation or continued work. A /codex:status command tracks running and recent jobs so you can manage concurrent tasks without losing track. The plugin requires a ChatGPT subscription or an OpenAI API key, and it's a telling signal: a competitor investing engineering effort to make its tool work well inside Claude Code means the ecosystem has grown large enough that ignoring it isn't an option.
github.com/openai/codex-plugin-cc
Anthropic's SDK Adds Persistent Agent Memory
Version 0.116.0 of the Anthropic Python SDK (released July 2nd) introduced opt-in support for agent memory via a new beta API header. Rather than carrying an ever-growing context window — which eventually hits token limits and drives up costs — or losing state between turns entirely, agents can now store intermediate conclusions and retrieve them on demand across separate API calls. It's a foundational primitive for stateful agent behavior: an agent that remembers what it's already ruled out, what it's waiting on, and what it decided three turns ago operates at a qualitatively different level than one starting fresh each call.
github.com/anthropics/anthropic-sdk-python
Managed Agents Gets Production-Ready Features in 0.115.0
The June 30 SDK release packed in substantial capabilities for teams running agents in production. Managed Agents gained event delta streaming for real-time progress updates, agent and deployment overrides, reverse pagination support, vault credential injection scoping (agents can use secrets without those secrets appearing in the prompt or logs), and webhook events for agent and deployment state changes. Individually these are table-stakes features for production systems; together they close the gap between "interesting prototype" and "something you can stake a real workflow on" for managed agent deployments.
github.com/anthropics/anthropic-sdk-python
Strix Surges Past 35k Stars as AI Security Testing Gains Momentum
The open-source AI penetration testing platform hit 35,000 GitHub stars today, adding nearly 2,800 in a single day. Strix coordinates multiple specialized AI agents to run penetration tests against codebases and live web applications, covers the OWASP Top 10, and validates every finding with a working proof-of-concept exploit rather than an unconfirmed alert. Unusually for a security tool, it integrates with GitHub Actions for CI/CD-gated security checks and supports automated patching workflows — the vision is continuous security testing woven into the development pipeline rather than a periodic engagement.