Code With Claude London 2026 Doubled Rate Limits And Shipped Five Agent Features — Here's What Actually Changes For Outsourcing Shops

Anthropic's "Code with Claude" London event ran 20–21 May 2026 with both keynote days streamed live. After the dust from yesterday's announcements, the headline that will get repeated on LinkedIn is "doubled rate limits." The headline that matters for anyone running an outsourcing shop is something quieter: Anthropic shipped a coherent stack for running AI agents inside a client's own infrastructure, for the first time. That is a different product than "we sell access to Claude."

1. The Doubled Rate Limits Are Real, But Aren't The Story

Claude Code rate limits doubled for Pro, Max, Team, and seat-based Enterprise plans, effective immediately. For a four-person agency running Claude Code as the daily driver, this kills the most common pain of 2025 — the mid-afternoon throttle that broke flow on Tuesdays and Wednesdays. It also quietly resets the per-seat economics: at the new ceiling, Claude Code is now plausibly cheaper per developer-hour than a junior engineer's time, which it was not three months ago. Re-run your team's tooling budget this week.

2. Self-Hosted Sandboxes (Public Beta) — The Real Unlock

The single most consequential announcement: self-hosted sandboxes for Claude Managed Agents are now in public beta. You can run the agent inside your own VPC, your own network perimeter, your own security controls, with the model still hosted at Anthropic but the execution environment under your roof. For agencies that build for regulated clients — healthcare, finance, government — this is the gate that has been blocking AI agent deployment for two years. Sandboxes mean you can finally say yes to "we'll deploy the AI tooling inside our environment" without spending six months convincing a security team.

3. MCP Tunnels (Research Preview) — Bringing Internal Tools Into Reach

MCP Tunnels let Claude reach internal MCP servers that live behind firewalls, without exposing them publicly. For agencies, this means an agent built on Claude can touch a client's on-prem ticketing system, internal ERP, or custom CMS without poking holes in their network. The work that used to be a three-month integration project is now a tunnel config and a one-line MCP registration. Expect this to compress the price of "AI agent for internal tools" engagements from six figures into the high-five-figure range over the next six months.

4. The Five Managed Agent Features Worth Knowing

Anthropic shipped five new managed-agent capabilities:

Dreaming — agents run background analysis during idle time, surfacing insights when the user returns. Useful for monitoring, alerting, and recurring report agents.
Outcomes — first-party telemetry that tracks what the agent actually accomplished, not just what it tried. Pricing conversations with clients change when you can show outcome graphs instead of token graphs.
Multi-Agent Orchestration — first-class support for one agent delegating to specialized sub-agents. Replaces the LangGraph-style glue most teams have been hand-rolling.
Claude Finance — ten pre-built finance agents (forecasting, AR/AP, expense triage, audit prep). Agencies who serve finance clients should price an "AI-augmented bookkeeping" line by next week.
Add-ins — agent components that drop into Office, Google Workspace, and Slack with no custom UI work. Three days of frontend dev that used to be required has just disappeared.

5. The Compute Story Is Why The Rate Limits Could Double

Underpinning everything: Anthropic announced a multi-year compute partnership with the SpaceX-operated Colossus 1 data center in Memphis, gaining access to over 220,000 Nvidia GPUs and 300 megawatts of power. This is the reason the rate limit doubled landed without an asterisk. For agencies, the takeaway is that "we'll need to throttle you mid-project" is now an unlikely scenario through end of 2026, which matters when you are quoting a six-month build with AI assistance baked in.

6. What An Outsourcing Shop Should Do This Week

Audit which client deals were stuck on "AI deployment inside their environment." Re-quote them with the self-hosted sandbox option. A meaningful fraction will unstick.
Pick one internal tool integration per client (ticketing, CRM, ERP) and propose a one-week MCP tunnel pilot. Cheap to deliver, hard for the client to ignore.
Stand up "Outcomes" dashboards for every existing Claude-powered project. The next pricing conversation gets easier when the dashboard already exists.
If you have a finance-vertical client, build a Claude Finance demo this weekend. The conversion rate on a working demo right now is unusually high.

My Take

For two years, the AI-agent conversation in outsourcing has been "we'll integrate the OpenAI API for you." That conversation does not survive this announcement. The real product shape going forward is "we'll deploy an AI agent stack inside your perimeter, with telemetry, with multi-agent orchestration, and with the integrations you already have." That is a much higher-margin product than reselling API access, and the agencies that pivot to it this quarter will print money for the rest of the year. The ones that stay on "we use Claude internally to ship faster" will compete on hourly rates with everybody else.

The doubled rate limits will get the headlines. The self-hosted sandboxes will pay the rent.