Skip to content

Deprecated & Outdated

Key Points

Things that were canonical in 2024 but should NOT be used in new code in 2026. The .NET AI ecosystem moved fast.

Semantic Kernel-only orchestration → Microsoft Agent Framework

Old (2024): Semantic Kernel as the canonical .NET AI orchestrator. SK Planners. SK ChatHistory. AutoGen for multi-agent.

New (2026): Microsoft Agent Framework (GA April 2026). Replaces SK orchestration + AutoGen. Built on Microsoft.Extensions.AI.

SK still useful for: existing apps; prompt template engine; some plugins. Don't START new projects with SK as the orchestrator. See SK vs Agent Framework.

SK Planners (Sequential, Stepwise, Handlebars) → FunctionChoiceBehavior.Auto

Old: SequentialPlanner, StepwisePlanner, HandlebarsPlanner generated multi-step plans via LLM and executed.

New: FunctionChoiceBehavior.Auto + iteratively-called LLM. The LLM decides when/which tools to call; runtime invokes; repeat. Modern reasoning models do this natively.

Planners are explicitly deprecated. See Deprecated Planners.

OpenAI SDK directly → Microsoft.Extensions.AI

Old: OpenAIClient from Azure.AI.OpenAI (or OpenAI package) used directly.

New: IChatClient abstraction. Wrap the vendor SDK; gain pipeline (caching, telemetry, function calling) + vendor portability.

// Old
var client = new OpenAIClient(apiKey);
var resp = await client.GetChatCompletionsAsync(...);

// New
IChatClient chat = new OpenAIClient(apiKey).AsChatClient("gpt-4o-mini")
    .AsBuilder().UseFunctionInvocation().UseLogging(...).Build();
var resp = await chat.GetResponseAsync("Hello");

Custom JSON tools → Function calling via attributes

Old: Hand-written JSON tool schemas; manual dispatch.

New: [Description] + reflection auto-generates schemas. UseFunctionInvocation() middleware auto-dispatches.

[Description("Get current weather for a city")]
async Task<Weather> GetWeather([Description("City name")] string city) => ...;

ChatHistory class → IList

Old: SK's ChatHistory class.

New: IList<ChatMessage> from Microsoft.Extensions.AI. Lighter; vendor-neutral.

BinaryFormatter for state → JSON / MessagePack

Same as elsewhere in .NET — BinaryFormatter removed in .NET 9. For agent state persistence, JSON or MessagePack.

Per-vendor embedding code → IEmbeddingGenerator

// Old
var emb = await openAiClient.GetEmbeddingsAsync(...);

// New
IEmbeddingGenerator<string, Embedding<float>> gen = ...;
var emb = await gen.GenerateAsync(["text"]);

Hand-rolled vector store → Microsoft.Extensions.VectorData

Old: Custom IVectorRepository with provider-specific code.

New: [VectorStoreKey], [VectorStoreVector] attributes; same DTO across Azure AI Search, Cosmos, Qdrant, pgvector. See Microsoft.Extensions.VectorData.

Old function calling formats → JSON Schema strict

OpenAI shifted to strict JSON Schema mode for function calling. Old "best-effort" function definitions superseded.

new ChatOptions { Tools = [AIFunctionFactory.Create(GetWeather)] };
// Schema auto-generated; strict mode default.

Single-vendor agents → MCP-pluggable

Old: Each agent has hard-coded tool implementations per vendor.

New: Tools served via MCP servers (any language). Agent becomes a thin orchestrator; tools are reusable across Claude, ChatGPT, Copilot, custom.

Long-context dump → RAG / structured retrieval

Old: Stuff entire knowledge base into prompt. Token-expensive; LLM accuracy degrades on long context.

New: Retrieve top-K relevant chunks; smaller, focused prompts. Even with 1M+ context windows (Claude, Gemini), retrieval improves quality + cost.

Hardcoded model strings → IChatClientFactory

Old: "gpt-4" literal everywhere.

New: Inject IChatClient via DI; model choice in config. Swap providers via configuration only.

CosmosDB without vector index → Cosmos NoSQL with VectorDistance

Old: Custom indexing of embeddings in Cosmos.

New: Native vector index in Cosmos NoSQL (.NET 9+ preview / GA depending on tier). SELECT TOP 10 ... ORDER BY VectorDistance(c.embedding, @e).

SignalR for streaming chat → IAsyncEnumerable + SSE

Old: SignalR hub for streaming LLM tokens.

New: IChatClient.GetStreamingResponseAsync() returns IAsyncEnumerable. Stream over SSE. Simpler.

await foreach (var update in chat.GetStreamingResponseAsync(prompt, ct))
    yield return update.Text ?? "";

"Prompt injection? Trust users" → Defense in depth

Old: Hope nothing bad happens.

New: - Spotlighting (mark user input). - Azure Content Safety Prompt Shields. - Output validation. - Don't give agents raw access to systems without confirmation.

Per-request model picking based on user → Routing as policy

Old: Free-form code branching on model.

New: Centralized policy (cheap → smart escalation; cost cap; tenant tier). Encapsulated routing.

Synchronous batch evals → CI eval pipelines

Old: Eval on demand.

New: Automated eval suite in CI. Track regressions. Tools: Ragas, Prompt Flow, custom.

What to AVOID starting fresh in 2026

  • ❌ SK as primary orchestrator.
  • ❌ SK Planners.
  • ❌ Direct OpenAI SDK (skip the abstraction).
  • ❌ Manual function calling JSON.
  • ❌ BinaryFormatter.
  • ❌ Hardcoded model names.
  • ❌ "Stuff everything in context" RAG anti-pattern.
  • ❌ No prompt injection defense.
  • ❌ No telemetry on AI calls.

What to CHOOSE in 2026

  • ✅ Microsoft.Extensions.AI everywhere.
  • ✅ Microsoft Agent Framework for orchestration.
  • ✅ MCP for tools.
  • ✅ Microsoft.Extensions.VectorData for vector stores.
  • ✅ Function calling via reflection + auto.
  • ✅ OTel GenAI for telemetry.
  • ✅ Eval pipelines in CI.

Cross-references