Claude Code Projects and Jev are shifting how we think about agent orchestration

The latest updates from Anthropic and TypeSafe show a clear move away from simple "chat + tools" toward persistent, coordinated agent runtimes. We're seeing a split between high-reasoning frontier models and fast, discriminative "System One" models that act as the routing layer for complex workflows.

Claude Code Projects brings multi-session orchestration to the cloud

Anthropic just rolled out Projects within Claude Code, which effectively moves "one conversation, many cloud threads" into a real product. Instead of a linear chat, a single conversation can now spawn parallel cloud sessions. These threads can pass context between each other and keep running even after you close the session.

Internally, Anthropic staff (Cat Wu and MikeyK) describe this as a coordinator abstraction. It uses aggregated status updates and evolving long-lived memory, all managed by a single controlling Claude instance. While these currently run in the cloud, local workflows are expected to follow. This is a significant step toward agents that can actually manage long-horizon work without losing the thread.

Google is standardizing agent infrastructure with new APIs

Google is pushing managed harnesses and secrets management to make agents more practical. They've updated Gemini managed agents with an Antigravity-based harness and two specific APIs:

• Credentials API: This keeps secrets out of the model context by using placeholders and trusted-domain egress proxying.
• Files API: This handles artifact movement and provides persistent sandboxes.

According to the release, these changes result in 22% higher cache hits and up to 30% lower costs. This trend toward persistent agents with scoped permissions and asynchronous execution is also appearing in Meta’s desktop Muse for Mac, Perplexity’s Computer, and Base44’s phone-calling Superagent.

Using Jev as a discriminative control flow primitive

TypeSafe’s Jev is gaining traction not as a chatbot, but as a "System One" classification tool for structured decision-making. Builders are treating it as an "AI if statement" or a routing layer to keep expensive frontier models focused on hard reasoning while Jev handles the high-frequency, low-latency tasks.

The community is using Jev for:

• LLM-as-judge and harness routing.
• Subagent creation and typed extraction.
• Reranking and instant routing.

Cloudflare has already integrated it via AI Gateway, and open reproductions like openjev-s (using Qwen3.6-35B-A3B with SGLang radix cache) have already appeared.

The debate over history compaction and KV caching

There is a tension between using classifiers like Jev for memory management and the actual economics of LLM caching. A common critique (notably from Theo) is that using Jev for aggressive line-by-line history compaction can be counterproductive.

The risk is twofold: dropping hidden reasoning payloads can degrade the performance of frontier models, and editing history often invalidates cached prefixes, which can actually increase costs. The real technical question isn't how to compact better, but whether future harnesses can eventually abstract away KV caching concerns entirely.