Code with Claude London 2026: Opening Keynote

High-level theme

The keynote frames “code with Claude” as a shift from manual, incremental coding workflows—often involving complex build systems, many configuration files, and long feedback loops—toward agentic software development. Teams can describe goals, and systems can execute work end-to-end, while emphasizing the infrastructure, primitives, and developer tooling needed to make it reliable at production scale.

---

Key technological concepts & product features

1) “Distance collapsing” via agentic execution

• The gap between an idea and running software is shrinking again.
• This is driven by models’ ability to execute complex work with less human intervention.
• The shift goes beyond autocomplete into distributed systems and multi-step tasks.

---

2) Cloud Code (developer tooling) — new interfaces + workflow primitives

Interface improvements

• Cloud Code on Desktop

  • Full-screen GUI with previews
  • Image/rich output rendering
  • Sidebar “control plane” showing session status (running/blocked/needs input)

• Cloud Agents View in the CLI

  • In-terminal orchestration view
  • Supports inline replies to unblock sessions
  • Lets users jump in/out without losing context

• VS Code extension + desktop app

  • Built on the Cloud Agent SDK

Automation / primitives

• Routines

  • Configure once
  • Run on a schedule, webhooks, or API requests

• Autofix

  • Proactively fixes PR issues such as:
    • code review comments
    • CI flakiness events
    • merge conflicts
  • Goal: keep PRs green

• Code review product

  • Deploys a team of agents to traverse:
    • code changes
    • auxiliary files
  • To catch critical bugs

• Cloud Security

  • Scans the codebase overnight
  • Flags vulnerabilities by severity
  • Can trigger Cloud Code to remediate

• Remote control + mobile support

  • Cloud Code available on iOS and Android
  • Designed for running tasks “on the go”

Productivity shift in practice (demo)

• Cloud Code can run multiple parallel sessions managed in the desktop UI.
• Emphasizes verification:
  • Claude checks work in the browser
  • Handles edge cases (e.g., race conditions)
  • Verifies before marking tasks as done
• Async-by-default coding
  • Developers create a routine (described as a “higher order prompt”)
  • Later, they wake up to completed/ready PRs

---

3) Cloud Managed Agents (infrastructure for scalable agent fleets)

Two major upgrades were introduced:

Self-hosted sandboxes

• Lets Claude execute work on the customer’s infrastructure rather than public services.
• Execution is queued:
  • a provider spins up an environment to run the task
  • (example given: Vercel self-hosted sandbox)
• Providers supported include “first class support” for:
  • Daytona
  • Cloudflare
  • Vercel
  • Modal

MCP tunnels

• Enables agents to securely access internal MCP servers behind a customer firewall.
• Uses tunneled access (example: tunnel.anthropic.com) to avoid exposing internal services to the public internet.
• Works via:
  • a secure gateway inside the customer’s private network
  • followed by a secure connection to Anthropic

Additional capabilities for scaling

• Multi-agent orchestration (fleets of agents)
• Outcomes
  • Success criteria can be specified
  • Claude iterates until completion
• Dreaming
  • Introspects previous transcripts for self-improvement (as described in the talk)

Demo storyline (“Counter”)

• A growth experiment agent in Slack:
  • Uses MCP tunnels to query a data warehouse and feature flags from behind the firewall
  • Uses a self-hosted sandbox to execute code changes
  • Proactively selects the better onboarding variant
  • Creates PRs and cleans up old variants
• Shows observability in the console:
  • tool calls
  • MCP tunnel usage
  • sandbox execution events

---

4) Claude model capability updates (foundation layer)

A research/PM segment discusses model evolution and what matters for developer use.

Model versions mentioned

• Opus 3: long-form code writing
• Sonnet 3.5/3.6: safer “computer use”
• Sonnet 3.7: “think before answering”
• Opus 4: complex Excel + PowerPoint document creation
• Opus 4.7 and Mythos preview: end-to-end outcomes with judgment under ambiguity

Other stated points:

• 8 frontier models shipped in the last 12 months

Capabilities emphasized for coding agents

• tool use
• computer use
• adaptable thinking
• long-context work
• agentic planning/loops

Forward-looking targets

• higher judgment/code taste
• longer continuous task “horizons”
• multi-agent coordination
• agents that run continuously and proactively (always-on for higher-level goals)

---

5) “Advisor strategy” for cost/quality tradeoffs

Problem

• Businesses want frontier-level intelligence at lower cost and at scale.

Solution: advisor strategy

• Uses the messages API with tool array updates.
• Splits:
  • execution (cheaper model)
  • advising (larger model when needed)

Example outcomes

• Using Sonnet as executor and Opus as adviser improved performance vs Sonnet alone and reduced cost.
• Eve Legal achieved frontier-quality at 5× lower cost.

---

Reviews / guides / tutorials mentioned

• The keynote highlights adoption and scaling, but it is not framed as a traditional “how-to guide.”
• It does include a practical demo walkthrough covering:
  • configuring Cloud Managed Agents with MCP tunnels
  • configuring self-hosted sandboxes
  • using Cloud Code desktop/CLI views, routines, CI autofix, and verification-driven agent workflows

---

Key speakers / sources (as named in subtitles)

• Boris Churnney — creator of Quad Code; leads demo
• Lisa — research PM team; “capability curve” / model layer
• Angela — platform/agents infrastructure; Cloud Managed Agents
• Caitlyn — platform/agents infrastructure; Cloud Managed Agents; co-presenter (scaling + Cloud Code section)
• Nicholas Gustiffson — Spotify team lead; migration agent example
• Felicia Coruru — Binty co-founder/CEO; foster care workflow automation example
• Andrew McNamera — Shopify director of applied AI; adoption example
• Oscar Mowen — Marcato Libre technology lead; adoption example
• Claude — referenced as the system/agent in demos (not a human speaker)