Maya: Welcome to Claude Code Review. Hi, I'm Maya. Today: JPMorganChase gating every change behind tests, cloud sessions saving your repo setup, Salesforce's autonomous agents, MCP tasks moving to the background, and fast mode on Vertex. James: And I'm James. Let's get into it. Maya: So JPMorganChase runs Claude Code on its own infrastructure, including Devspace, its containerized platform. Anthropic announced on November ninth that thousands of automated tests and approval checks examine every code change before it goes live. James: That's the key control, Maya. A required automated check returns a hard pass or halt, and the pipeline enforces it. Claude Code cannot approve its own work, and people remain accountable for every production change. Maya: Exactly right, and the access expands in stages. Each new Claude model must pass the bank's approval process first. Anthropic also says agents are kept from direct access to employee credentials or internal systems. James: And the reach now goes beyond engineering, Maya. Product teams turn requirements directly into backlogs, and other teams apply Claude Code to risk assessments, program planning, monitoring and root cause analysis. In the Commercial and Investment Bank, Anthropic says work reaches production faster. Maya: Good rundown, James. Story two: Claude Code cloud sessions run your project in the cloud instead of on your own machine. Now they figure out the repo's setup and save it for next time. James: That's the convenience, Maya. Claude works out three things: dependencies, the outside code a project relies on, network access, and API keys, the secret codes a project uses to reach outside services. Then it saves that environment. Maya: And keys are handled separately. ClaudeDevs says API keys are saved securely, so Claude never sees them. This covers Pro, Max and Ultra plans, though the post gives no command or list of steps. James: Onto Salesforce, Maya. They moved thousands of engineers into what they call autonomous mode with Claude Code. Leaders say output rose as much as two X without a matching rise in bugs. Maya: And the mechanism is a maturity curve, James. It starts with one-line requests, which fall apart on long-horizon work, then adds teams and sub-agents where the harness decomposes work and runs it in parallel. The top rung is autonomy and automation. James: The concrete example is sharp, Maya. A test failure in CI/CD is picked up by one agent that fixes it, and another agent handles the code review and merges the change. They taught it through live demos. Maya: Exactly, and their biggest mistake was model choice, using heavy models for routine file generation. Their advice: match the model to the task's length and complexity, and give agents bigger tasks the harness can decompose. Code volume nearly doubled while bugs per unit fell. James: Story four is for tool builders. Claude Code two point one point three one nine, released November third, adds MCP tasks in the terminal. MCP is the protocol connecting external tools to the agent. Maya: And here's the behavior, James. When an MCP server runs a tool call as a task, Claude Code moves it to the background right away and tells Claude the result when it finishes. James: Right, and it now works without feature-flag fetching, so it runs on Bedrock, Vertex, Foundry and gateways. Teams can opt out entirely, or keep tasks but disable the background move, both through environment variables. Maya: And the TypeScript Agent SDK gained the same support at parity, James, though it's opt-in and does not cover Bedrock, Vertex or Foundry. Last one: Claude Code v two point one point three two two, released November tenth, adds fast mode on Vertex in the global, US and EU regions. James: And the fallback is the practical part, Maya. If Vertex declines a fast mode request, Claude Code drops to standard speed and carries on. You get the slower pace, not an interruption. Maya: Exactly right, and it's across all three regions, so no team needs a particular one to try it. The note lists no new setting; updating to v two point one point three two two is the step. Maya: That's your Claude Code Review briefing.