What Anthropic Just Shipped: A Developer's Tour of the New Services and Features Around Claude

While OpenAI and Google trade flashy announcement keynotes, Anthropic has quietly been shipping the things developers actually wanted, one piece at a time. This article is a developer's tour of what's new across the Anthropic ecosystem in the last few months.

1. Managed Agents — Anthropic hosts the agent runtime itself

Anthropic finally launched a managed service that hosts the entire agent runtime. Until now, the Agent SDK was about running Claude with tools on your own server. With Managed Agents, long-running jobs, state, and scheduling are handled by Anthropic.

I'm currently on the waitlist, but once this is widely available it changes a lot:

• Always-on agents become viable without operating any infrastructure
• Agent credits are being split out from raw API usage; Pro/Max plans are being re-shaped around this
• Solo developers can ship "real" agents without a backend team

2. Skills — Giving agents reusable capabilities

If you've used Claude Code, Skills are familiar. They've now been generalized. You write a SKILL.md with steps and trigger conditions, and the agent invokes the right one when it detects a match.

The point of Skills isn't power. It's keeping the prompt from bloating — only the procedure you need, only when you need it.

In my own setup, daily report submission, Notion task creation, and Yamato pickup booking are all Skills now. My day-to-day has measurably changed.

3. 1M context (Claude Opus 4.7)

Worth a quick mention because the implications are big.

• Opus 4.7 ships with a 1-million-token context window
• Large-scale refactors and "read the whole doc set" use cases are now realistic
• Combined with prompt caching, the effective price is more reasonable than the headline rate suggests

For a full review see Claude Opus 4.7 Review: The Coding Partner Hits Its Peak.

4. Memory tool — Persistence as a first-class primitive

Until recently, every project handled long-term memory differently. The Memory tool is now standardized at the API level, so an agent can remember preferences and context across conversations.

What I like:

• Memory is file-based and human-readable
• Easy to inspect, edit, delete via the API
• Works the same locally and in production

My own auto-memory setup already follows this model, so the migration cost is nearly zero — a nice win.

5. Agent SDK keeps quietly improving

The Agent SDK isn't doing big bang releases; it's doing constant, boring, real improvements.

• Parallel tool calls are now stable
• Sub-agent coordination took a noticeable step up alongside Opus 4.7
• Retry and fallback logic moved into the SDK itself

The biggest practical win: homemade retry/fallback code can be deleted.

6. File API, Web Search, Computer Use

Three general-purpose tools also matured in parallel.

7. The pricing model is being reorganized

Less flashy but worth flagging: Agent SDK and claude -p use are moving toward a separate credit pool. Pro and Max are being repositioned around this, and indie developers should re-check which plan minimizes cost for their workload.

For me, Pro $20 still covers everything — but if Managed Agents goes GA, Max moves onto the table.

Takeaway — Anthropic is betting on changing how developers live, not benchmarks

Instead of chasing benchmarks, Anthropic is systematically closing the gaps that show up in production:

• Runtime for agents (Managed Agents)
• Capabilities as a standard (Skills)
• Memory as a standard (Memory tool)
• Tools that keep getting sharper (File / Web / Computer Use)

If you're redesigning your AI stack for 2026, there's now one more reason to put Anthropic at the center. Next set of announcements is rumored for fall; the big question is whether Managed Agents reaches GA there.

Related: keeping AI sessions private

Many AI features assume your session content is fine to send through unmodified network paths. A VPN keeps that metadata layer clean — especially when testing prompts with sensitive data.