✦ Claude Code Review — Multi-Agent PR Analysis

Launched on 9 March 2026 and now in research preview for Claude Team and Enterprise customers, Code Review for Claude Code dispatches a team of specialised agents that analyse a pull request in parallel — each agent focused on a different class of issue. When they finish, they post a consolidated comment on the PR with findings and suggested fixes. Anthropic's own engineers using the feature internally reported a 200 % increase in output.

How it works

• One lead agent decomposes the PR into review domains (security, logic, style, test coverage, etc.).
• Specialist agents work in parallel, each owning their domain and maintaining their own context window.
• Agents communicate findings to the lead, which synthesises a single GitHub comment.
• The comment highlights issues and proposes targeted fixes for human review and approval.

Cost & timing expectations

• Typical cost: $15–$25 per review.
• Typical completion time: ~20 minutes.
• Cost scales with PR size and the number of parallel agents dispatched.

✦ Agent Teams in Claude Code — Parallel Sub-Agents

Shipped alongside Opus 4.6 in February 2026, Agent Teams is an experimental Claude Code feature that lets a single session spin up a coordinated team of Claude Code sub-agents. Rather than one session working sequentially through a large task, the lead session can split work across teammates that each own a piece — with their own independent context windows — and then coordinate results back to the lead.

Architecture pattern

• One session acts as the team lead — it plans, delegates, and integrates results.
• Teammate sessions work independently in parallel, reducing wall-clock time on large, parallelisable tasks (e.g. refactoring many modules simultaneously).
• Teammates can communicate directly with each other, not just back to the lead.

Enabling the feature

Agent Teams are disabled by default. Enable them by adding the flag to your Claude Code settings or environment:

# In settings.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "true"
  }
}

✦ Structured Outputs GA & Data Residency Controls

Two significant API production-readiness upgrades landed in early 2026. Structured Outputs — which guarantee Claude returns valid JSON conforming to a developer-supplied schema — are now generally available across all current Claude models with no beta header required. Separately, Data Residency Controls let teams constrain where model inference actually runs on a per-request basis, answering a common compliance requirement.

Structured Outputs — what's new at GA

• No anthropic-beta header needed — use it in production today.
• Expanded schema support: richer JSON Schema constructs are now honoured (e.g. anyOf, recursive schemas).
• Improved grammar compilation latency — time-to-first-token is meaningfully lower than the beta.
• Available on Opus 4.6, Sonnet 4.6, Sonnet 4.5, Opus 4.5, and Haiku 4.5.

response = client.messages.create(
  model="claude-sonnet-4-6",
  max_tokens=1024,
  messages=[{"role": "user", "content": "Extract the invoice fields."}],
  tools=[{
    "name": "extract_invoice",
    "input_schema": {          # ← your JSON schema here
      "type": "object",
      "properties": {
        "vendor":  {"type": "string"},
        "amount":  {"type": "number"},
        "due_date":{"type": "string", "format": "date"}
      },
      "required": ["vendor", "amount"]
    }
  }],
  tool_choice={"type": "tool", "name": "extract_invoice"}
)

Data Residency — inference_geo

The new inference_geo API parameter (also settable as a workspace default) controls which region processes a request. Currently supported value: "us" for US-only inference. This is useful for teams with data-sovereignty requirements that prohibit processing outside a specific jurisdiction.