🧭 Claude Founder House Berlin: Anthropic's First European Developer Day

Anthropic ran its inaugural Claude Founder House Berlin today — a full-day live event in Berlin (streamed from 09:00 CEST) aimed at founders and developers building European companies on Claude. The format mirrors Anthropic's Founder House series in San Francisco and New York, adapted for the European context: a morning Founder cohort combining keynote sessions with peer programming, and an afternoon Builder cohort focused on hands-on construction with live Anthropic engineers on hand.

What the morning keynotes covered

• GDPR-native Claude architectures: Anthropic's EU-based engineering lead walked through how to build Claude integrations that never send EU personal data outside Europe, using the inference_geo: "eu" parameter released in February and the new EU data-residency endpoint (api.eu.anthropic.com). The session included a live diff of a before/after integration to make the migration concrete.
• Agentic product patterns for regulated industries: A founder panel (three Berlin-based companies building in fintech, healthcare, and legal) discussed what it actually takes to deploy Claude agents under EU regulations — MiFID II data-handling constraints, medical-device software rules, and German legal-tech ethics guidelines. The consensus: start with the most constrained use case in your category, get that right, then expand.
• EU AI Act tier mapping: Anthropic published a short reference guide mapping Claude's built-in safety features to the EU AI Act's risk tiers. Applications built on Claude in the "high-risk" categories (HR, credit, education, biometrics) must document their conformity assessment; the guide explains what Anthropic supplies and what the application developer must add.

Afternoon Builder cohort highlights

The afternoon session included three hours of structured building time, with Anthropic engineers rotating between tables. Common patterns that emerged:

• Tool-call caching for GDPR-sensitive pipelines: Several teams discovered that ephemeral prompt caching (cache lifetime set to zero, so nothing persists past the session) lets them get most of the latency benefit of caching without triggering data-residency concerns for sensitive fields.
• Structured outputs for regulatory evidence trails: Using response_format: json_schema with a schema that includes a reasoning_trace field forces Claude to produce a machine-readable audit log alongside its output — useful for demonstrating compliance in regulated workflows.

🧭 G7 AI Communiqué Published: What the Risk-Tiered Deployment Framework Means for Claude API Developers

Following Dario Amodei's briefing to G7 leaders yesterday in Evian, the full G7 AI Governance Communiqué was published this morning. The central mechanism is "risk-tiered deployment authorisation" — a three-tier classification system that determines how much pre-deployment review a model or application requires before it can be commercially deployed. The framework is non-binding in its current form but is described as a template for domestic legislation.

The three tiers

• Tier 1 — General AI applications: Standard consumer and business tools. No new pre-deployment requirements beyond what already exists in each jurisdiction. The vast majority of Claude API use cases fall here. Developers building chat assistants, document summarisers, code review tools, and similar productivity applications are in Tier 1.
• Tier 2 — High-impact AI applications: AI used in high-risk decision contexts (credit, employment, healthcare triage, legal advice). Requires documented conformity assessment and an ongoing incident-reporting obligation. Anthropic's EU AI Act guide (published today) maps directly to this tier for EU-based deployments.
• Tier 3 — Frontier model training and deployment: Applies only to model developers training above the 10²⁶ FLOPs threshold proposed in Amodei's briefing. Requires a pre-deployment capability evaluation by an independent third party before commercial release. This tier does not apply to Claude API users — it applies to Anthropic itself (and its peers).

What the communiqué does not do

Notably absent: any binding international coordination mechanism on compute thresholds, which Amodei had specifically requested. G7 members agreed on the principle of aligning definitions of "frontier training run" but deferred the specific FLOPs number to a subsequent technical working group. The working group is expected to report back at the G7 AI ministerial in October 2026.

🧭 Claude Code Gains Nested Sub-Agent Spawning — Agents That Create Their Own Agents

A significant Claude Code release shipped overnight, introducing nested sub-agent spawning: the ability for a sub-agent to spawn its own child agents. Until now, Claude Code's multi-agent mode was strictly two-level — an orchestrator could spawn sub-agents, but sub-agents could not themselves spawn further agents. That constraint is now lifted. The result is true hierarchical multi-agent architectures: an orchestrator can delegate to a project manager agent, which can itself spin up specialist agents for individual sub-tasks, with results rolled up back through the tree.

How to use it

Nested spawning is enabled by default when --agents is passed to Claude Code. Sub-agents that want to spawn children use the same spawn_subagent tool call their parent used. The depth limit is configurable:

# Default: depth limit of 3 (orchestrator + 2 levels of nesting)
claude --agents "Refactor the payments module"

# Increase depth for deeply hierarchical tasks
claude --agents --max-agent-depth 5 "Audit the entire codebase for security issues"

# Disable nesting (revert to two-level-only)
claude --agents --max-agent-depth 1 "Run all the tests"

Also in this release

• /cd command: Change the session's working directory without rebuilding the prompt cache. Previously, moving to a different directory required starting a new session, losing the context of everything discussed so far. /cd ../sibling-service now switches the working tree in place.
• Mid-stream disconnect recovery: Claude Code now preserves the partial response when a connection drops mid-generation. Previously, a mid-stream drop produced an error and discarded the partial output. Now the partial response is saved; on reconnect, you see what was generated before the drop and can continue from that point.
• WSL2 / Windows Terminal fixes: Mouse-wheel scrolling through long outputs in WSL2 and VS Code's integrated terminal was broken in the June 11 release. This is fixed.