[bot] Anthropic Claude Managed Agents (CMA) session execution not instrumented

[braintrust-bot]

Summary

The Anthropic Ruby SDK (v1.31.0+) supports Claude Managed Agents (CMA) — a first-party agentic execution surface where users send messages to agent sessions and stream back generative output including tool use, text, and status events. The key execution methods (client.beta.sessions.events.stream_events and client.beta.sessions.events.send_) are completely uninstrumented by this SDK.

CMA is an agentic run/tracing surface: a user sends a prompt, and the agent autonomously generates text, invokes tools (bash, file I/O, web search), and streams results back. This is functionally similar to multi-turn chat completions with tool calling, but runs inside Anthropic-managed containers.

What is missing

Generative execution surfaces (not CRUD)

• client.beta.sessions.events.stream_events(session_id) — Opens an SSE stream that yields agent execution events: agent.message (text generation), agent.tool_use (tool invocations), agent.tool_result, and session.status_idle (completion). This is the primary consumption surface for agent output.

• client.beta.sessions.events.send_(session_id, events:) — Sends user messages that trigger agent generation. Each send_ call initiates a new turn of agentic execution.

What a span should capture

• Input: the user message content sent via send_
• Output: aggregated agent messages and tool use events from the stream
• Metadata: agent ID, session ID, environment ID, model, tools, provider (anthropic), endpoint (/v1/sessions/{id}/events), api_version: "beta"
• Metrics: token usage (if available in stream metadata events), time_to_first_token, tool execution count

Pattern precedent

This SDK already instruments the Anthropic beta messages API (client.beta.messages) with defensive error handling for beta API changes (see lib/braintrust/contrib/anthropic/instrumentation/beta_messages.rb). A similar approach would work for CMA session events.

Braintrust docs status

not_found — Braintrust's Anthropic integration docs at https://www.braintrust.dev/docs/integrations/ai-providers/anthropic document messages API tracing only. No mention of Claude Managed Agents, sessions, or agentic execution instrumentation.

Upstream sources

• Anthropic CMA quickstart with Ruby examples: https://platform.claude.com/docs/en/managed-agents/quickstart
• Anthropic CMA overview: https://platform.claude.com/docs/en/managed-agents/overview
• Anthropic Ruby SDK releases: https://github.com/anthropics/anthropic-sdk-ruby/releases — CMA support added in v1.31.0 (April 8, 2026), CMA Memory added in v1.36.0 (April 23, 2026)
• Ruby API surface: client.beta.sessions.events.stream_events(session_id) and client.beta.sessions.events.send_(session_id, events:)

Local repo files inspected

• lib/braintrust/contrib/anthropic/patcher.rb — defines MessagesPatcher and BetaMessagesPatcher only; no session/agent patchers
• lib/braintrust/contrib/anthropic/integration.rb — registers [MessagesPatcher, BetaMessagesPatcher]; no CMA patchers
• lib/braintrust/contrib/anthropic/instrumentation/ — contains messages.rb, beta_messages.rb, common.rb; no session or agent files
• Grep for session, agent, cma, managed across lib/braintrust/ returns zero matches related to CMA