ai-agents

Keep Claude Managed Agents, skills, environments and memory stores in git and reconcile them with `ant apply` (anthropic-cli 1.30.0+, current 2.9.0). The file layout for each kind, what claude-lock.json records and why you must commit it, how drift, --force and --prune behave, why memory stores attach through deployments rather than agents, and a GitHub Actions workflow that applies on merge.

The Anthropic Python and TypeScript SDKs now emit OpenTelemetry spans on their own: register a tracer provider before the first call and every messages.create becomes an anthropic.messages.create CLIENT span with gen_ai.* token, model and stop-reason attributes. Content mode, conversation IDs, user.id baggage, linking to the Claude Platform's server-side spans, streaming and tool runner behavior, and the env-var override trap, checked against anthropic 2.8.0.

Microsoft Agent Framework for .NET 1.28.0 removes A2AServerRegistrationOptions.AgentRunMode, so hosted A2A agents always answer with a message, and replaces the id/name/description overloads on A2AAgent and AsAIAgent with A2AAgentOptions. Here is what breaks and how to migrate.

Send diagnostics.previous_message_id on every Messages API call and Claude tells you which part of the prompt broke the cache: model_changed, system_changed, tools_changed or messages_changed, plus an estimate of the tokens you paid to rewrite. Request and response shapes, the pending and not-found states, streaming, a logging helper for anthropic 2.7.0, and the limits.

Stop editing the tools array between turns. Append a role: system message with tool_addition and tool_removal blocks (beta inline-tools-2026-09-15) so a Claude agent gains, loses or re-versions tools while the cached prefix and preserved thinking stay valid. Placement rules, defer_loading, schema changes, error codes and model support.

Use the Claude API's on-demand compaction parameter (beta compact-2026-09-04) to summarize a tool-calling agent loop before it outgrows the 1M window: when to trigger it, where to cut so tool calls and thinking stay valid, how to write the summarization instructions, and the stop reasons, errors and billing fields that trip people up.

ModelContextProtocol 2.4.0 ships ModelContextProtocol.Extensions.Skills, an experimental package that implements the SEP-2640 Skills extension: one call to publish a folder of SKILL.md directories, and client helpers that verify every file against a SHA-256 manifest.

Copilot Fusion, generally available since October 28, 2026, is a runtime model router that picks a model and an execution pattern (Single, Cascade, or Critique) per prompt. Here is how the three patterns work, what Fusion Standard, High, and Max cost you in AI credits, how to lock the effort level with managed-settings.json, and the gotchas the docs bury.

ACP (Agent Client Protocol, v1) connects an editor to a coding agent; MCP (spec 2026-07-28) connects an agent to tools. They sit on opposite sides of the agent and are used together: ACP's session/new even carries the MCP servers the agent should start. Here is the feature matrix, the wire flow, and a working agent and client checked against @agentclientprotocol/sdk 2.4.0.

MXC is Microsoft's MIT-licensed SDK for running untrusted code, including model output, plugins and tools, under one deny-by-default policy that it maps to Seatbelt on macOS, bubblewrap on Linux and ProcessContainer on Windows. Here is what MXC SDK 1.0.1 and schema 1.0.0 actually define, how Copilot CLI hands it a policy per process, and what stays outside the box.