Anthropic releases new API features for building agents: a code execution tool, an MCP connector, a Files API, and extended prompt caching, all in public beta
Today, we're announcing four new capabilities on the Anthropic API that enable developers to build more powerful AI agents …
Anthropic
Context & Ripple Effects
This expands the developer foundation Anthropic began with Claude tool use in its Messages API, moving from model-level actions toward a broader set of services for agent applications.
In later coverage, Anthropic packages more of that stack into Managed Agents with an agent harness and deployment tools. The sequence matters because it shifts more agent infrastructure from application builders to the API provider.
First-order effects
Developers can combine code execution, file handling, external-tool connectivity and longer-lived prompt context within Anthropic's API, reducing the amount of surrounding infrastructure they must assemble themselves.
The MCP connector makes integrations more central to agent design, while MCP's cited token overhead makes context-management features directly relevant to the cost and practicality of those integrations.
Second-order effects
Agent developers and infrastructure vendors face a clearer build-versus-buy choice as Anthropic supplies more of the execution, data-access and context layers alongside the model.
Connecting agents to more files and tools increases the importance of defining an agent's memory and review process as well as the permissions governing what it can access and execute.
Third-order effects
If providers continue bundling agent primitives into their APIs, competition may move beyond model quality toward the completeness of each provider's agent runtime and integration ecosystem.
MCP's potential as a common connector layer could reduce bespoke integrations, but its token cost leaves room for optimization and may shape which agent workflows are economical at scale.
The trend: Foundation-model providers are evolving from model APIs into integrated platforms for building, connecting and operating AI agents.
Lots to be excited about from Anthropic today, but one I'm particularly jazzed about is that they did a bunch of welfare testing on Opus 4. I think current models are most likely not meaningfully sentient, but if they are it's of huge moral importance. See sec 5 here: www.anthr…
Just ran Claude 4 Opus on a project outline that I had already revised after getting feedback from o3 — and it gave me useful feedback on risks and missed opportunities that had eluded both me and o3. — Not sure whether to be happy about this or miserable. [embedded post]
I had early access to Claude Opus 4, and I've been very impressed by its ability to reason between multiple tool executions. — Here's my story on how I tested it, and the complex MCP workflows I built with it: — https://www.macstories.net/... [image]
a large part of why MCP has been so successful is due to how open the Anthropic team has been to collaboration. it's been great working with them, and congrats to their team on the launch of Claude 4 today!
Big MCP news today from Anthropic. Claude x Zapier just got a serious upgrade 👇 Claude 3.7 Sonnet can now take real-world actions (like sending a Slack message, updating a CRM record, or editing a Google Sheet) across nearly 8,000+ @zapier-connected apps. This is a BIG leap [vide…
Big day for the folks at @AnthropicAI! Super cool to see @getsentry pop up in their announcement today around all the work done on the Sentry MCP server. It's been a fun project to work on and see start to grow. Congrats on the announcements Anthropic! [image]
PayPal's remote #MCP server now supports the @AnthropicAI API, alongside our new #OpenAI integration. Bring together the power of #Anthropic plus PayPal's currently enabled #commerce capabilities today to build smarter, faster, and safer! Watch a short video. 👇 #AgenticAI [video]
Today, we're announcing four new capabilities on the Anthropic API to help developers build more powerful AI agents. A code execution tool, MCP connector, Files API, and extended prompt caching: [video]