Anthropic, Google, and OpenAI advance agent orchestration with persistent sessions, coordinator UIs, and domain-specific products

Anthropic’s Claude Code Projects introduces a coordinator abstraction that allows a single conversation to spawn parallel cloud threads, pass context between sessions, and continue running after the user disengages. Internal staff describe it as a higher-level coordinator with evolving long-lived memory and aggregated status updates surfaced via a single controlling Claude interface. The company notes that threads currently run in the cloud, with local workflows planned for a future release.

Google updated its Gemini managed agents with a new Antigravity-based harness and two APIs: a Credentials API that keeps secrets out of model context via placeholders and trusted-domain egress proxying, and a Files API for artifact movement and persistent sandboxes. Google claims the update delivers up to 30% lower costs and 22% higher cache hits. The release aligns with a broader industry push toward persistent agents with scoped permissions, user-specific context, and asynchronous execution as the default user experience rather than an optional add-on.

OpenAI launched Astra for Law, a domain-specific product that bundles 26 partner-built plugins and 47 community plugins. Initial rollout occurs through Trusted Access in ChatGPT and Codex, with API access planned for a later date. OpenAI reports that Astra for Law outperforms generic GPT-6 Astra with web search on its internal legal benchmark across price points. The packaging reflects a strategy of turning frontier reasoning capabilities into verticalized products with maintained configurations, tools, and safety defaults.

TypeSafe’s Jev is emerging as a fast, low-cost structured-decision primitive used for routing, LLM-as-judge, harness routing, subagent creation, and typed extraction. Community examples include open reproductions such as openjev-s with Qwen3.6-35B-A3B and SGLang radix cache, as well as browser demos. The technical thesis frames Jev as a discriminative control-flow layer that replaces prompts with explicit classification logic, enabling high-frequency decisions to be handled by small, low-latency models while reserving expensive frontier models for harder reasoning steps.

Google’s DeepMind published Stellar Colosseum, a model-agnostic many-agent harness designed for mathematics and theoretical computer science. The harness separates strategy, decomposition, subproblem solving, and verification, and reports claimed results of Codeforces 4263 and 71.0% on TCS-Bench. The work exemplifies a broader trend toward explicit memory structures, decomposition patterns, and reproducible benchmarks in multi-agent research.

Anthropic also released internal metrics tracking AI-driven R&D, including the share of model R&D tasks led by Claude rising from 1% to 26% in approximately six months, more than 90% of model R&D work involving Claude collaboration or leadership, and roughly 30,000 internal agents active. While these figures should be interpreted cautiously, they provide one of the few public glimpses into internal automation practices at an AI lab.

Benchmark practices are evolving beyond raw scores toward methodological scrutiny. Epoch launched Benchmark Reviews with 15 audits labeled Verified, Flawed, or insufficiently documented, and Vals introduced Vibe Code Bench 1-100 to measure iterative modification robustness rather than first-pass success. These efforts reflect growing attention to whether benchmarks themselves are fit for purpose.

Security remains a salient concern as agentic systems scale. Reports described a Claude-assisted compromise of OpenAI-connected accounts and internal repository access, achieved by chaining an image-upload bug, ChatGPT/Codex account takeover, and access to OpenAI-connected services, culminating in a pull request to OpenAI’s internal monorepo.