ai-agents

On November 3, 2026 the .NET team showed agent skills shipping inside NuGet packages as skills/<name>/SKILL.md. Here is the layout, how restore exposes it through project.assets.json, the two tools that read it today, and what is still missing.

The MCP Tasks extension (io.modelcontextprotocol/tasks, spec 2026-07-28, SEP-2663) lets a server answer tools/call with a durable task handle instead of a result. The client polls tasks/get, answers mid-flight questions with tasks/update, and can cancel or reconnect. Here is the wire protocol, what changed from the experimental 2025-11-25 version, when a task is worth it, and a working server and client checked against @modelcontextprotocol/ext-tasks 0.2.2.

Claude Tag is Anthropic's shared Slack agent: one org-owned identity per channel scope, with channel memory, admin-granted connectors, a GitHub App for draft PRs, and a monthly spend limit. Here is what it is, how setup and pairing work, how it differs from the old Claude in Slack and Claude Code in Slack apps, and the data-retention and billing details to check before you turn it on.

Microsoft Agent Framework for .NET 1.27.0 makes DefaultHttpRequestHandler throw HttpRequestException when its client-provider callback returns a client for a redirect destination. Here is why that hole existed, what breaks, and how to rewrite the callback.

A coding agent loses its context every session, so a project that runs for days needs its memory in the repository. Make the coordinator a small script and a JSON ledger, run every worker as a fresh `claude -p` session in its own git worktree with a hard scope and budget, and have the coordinator verify and merge the work itself. Tested with a stub worker; flags pinned to Claude Code 2.1.310.

The Anthropic SDK retries a failed call twice in about 1.2 seconds, which is fine for a chat request and wrong for a 200-turn agent run. Classify every failure into retry-after, capacity, transient or terminal, retry the model turn (never the tool call), share one pause gate across workers, catch the mid-stream overloaded_error that arrives as a 200, and checkpoint so a dead run resumes. Tested against anthropic 2.3.0 for Python.

A guard hook that crashes or times out lets the dangerous command through in most coding agents. Fail closed for pre-action guards, fail open for everything else, and make the script fail closed by itself because Copilot, Codex and Gemini CLI treat a timeout as allow. Compared on Claude Code 2.1.295+, Cursor hooks, Copilot CLI 1.0.105, Codex and Gemini CLI, with a portable guard script.

Moving an agent from claude-sonnet-4-6 or claude-opus-4-7 to claude-sonnet-5-5 or claude-opus-5-5 is not just a model ID swap. Thinking is always on, thinking-off and forced tool_choice now return 400, the effort defaults moved, Sonnet counts about 30% more tokens, and Claude Code ignores your old effortLevel on Opus 5.5. Here is the full checklist with before/after code.

If a specialist agent is just a model loop wrapped around a few read-only calls into your own services, collapse it into a skill served over MCP: a SKILL.md plus typed tools that the parent model calls directly. Keep it an A2A agent when it needs its own model, private context, a long multi-step workflow or an independent lifecycle. Here is the decision matrix, the measured cost of each path, and the Agent Framework 1.17 wiring.

Run the orchestrator on claude-opus-5-5, give implementer and reviewer subagents claude-sonnet-5-5, and send search, triage and summarizing work to claude-haiku-5-5 at medium effort. Here is the role matrix, the cost math per subagent run, Haiku 5.5's 100k-token price cliff, and the Claude Code model-resolution rules that silently override your choice.