Skip to main content
Status: Accepted (March 2026) Context: As LLM capabilities evolve rapidly, we need a framework to decide where to invest engineering effort and where to hold still.

Decision

We classify every feature by its relationship to LLM progress, and allocate effort accordingly.
CategoryStrategyInvestmentExamples
OrthogonalModels getting smarter doesn’t diminish these — pure engineering/integration problemsFull investmentConnectors, credentials, OAuth, audit, RBAC, security, deployment
TailwindModels improving makes these better, not redundant — symbiotic relationshipInvest (benefits compound)AI Connector Builder (smarter model = higher quality connector output)
FrozenAlready shipped, working well — but models are absorbing these capabilitiesMaintain only, no new featuresReAct loop, DAG planning, RAG pipeline, Memory, Grounded Generation
ConsiderProviders building natively at the platform level — high risk of redundancyDeferred indefinitelyMulti-Agent orchestration, semantic memory, memory lifecycle
Rule of thumb: If a feature solves “how to make the model smarter”, it’s being absorbed. If it solves “how to connect the model to the real world safely”, it’s orthogonal.

Analysis

Why Connector Platform is fully orthogonal

Models will never natively:
  • Store and encrypt API credentials (AES-GCM)
  • Manage OAuth flows (authorization page → callback → refresh token)
  • Connect to a client’s Kingdee/金蝶 ERP database
  • Push notifications to Lark/飞书 or WeCom/企微
  • Enforce RBAC on who can use which connector
  • Log every tool call for compliance auditing
These are engineering problems, not intelligence problems. A model 10x smarter still can’t do these things without infrastructure.

Why AI Connector Builder is “tailwind” not “being absorbed”

The Builder Agent uses model intelligence to create managed, persistent Connector entities — stored in DB, reusable across agents, with credential management and audit trails. The model’s improving API understanding makes the Builder produce better connectors, not makes the Builder unnecessary. Analogy: Cursor uses Claude to write code. Claude getting smarter makes Cursor better, not redundant, because Cursor provides engineering value (project management, file organization, version control) that the model doesn’t replace.

Why v0.1–v0.5 features are “frozen”

FeatureWhat’s happening in the industry
ReAct loopModels have native tool calling (OpenAI, Anthropic). The external reasoning loop adds less value as models internalize it.
DAG PlanningModel reasoning capabilities are improving rapidly. Complex task decomposition that needed external planners is becoming a single-shot capability.
Memory managementContext windows are growing fast (Gemini 2M+, Claude 200K+). The need for external window management, summarization, and compaction is shrinking.
RAG pipelineProviders are building retrieval into their platforms (OpenAI file_search, Google NotebookLM, Gemini Search Grounding). For public knowledge, the traditional chunk-embed-retrieve pipeline is being replaced.
Grounded GenerationModels are getting better at citing sources natively. The 5-stage grounding pipeline we built adds diminishing value.
These features are not bad — they shipped, they work, they make the product functional today. The decision is simply to stop adding to them and redirect effort.

Why Multi-Agent Orchestration was deferred

LLM providers are building orchestration natively:
  • OpenAI Swarm: Multi-agent framework with handoff protocols
  • Anthropic Claude Code Teams: Leader/Worker agent pools with task graphs
  • Google A2A (Agent-to-Agent): Inter-agent communication protocol
Building a competing orchestration layer would mean racing against first-party implementations with deeper model integration. This is not a sustainable differentiator.

Why Semantic Memory and Memory Lifecycle were deferred

  • Context windows are growing rapidly, reducing the need for cross-session memory retrieval
  • Providers are adding native memory features (ChatGPT Memory, Claude Projects)
  • The engineering cost of building a reliable memory system (TTL, importance scoring, semantic retrieval) is high relative to the shrinking gap it fills

Feature-Level Classification

Orthogonal (v0.6+)

FeatureVersionWhy orthogonal
Connector entity + CRUDv0.6.1Enterprise integration, pure engineering
Per-user credentials (AES-GCM)v0.6.2Security infrastructure
Confirmation Gatev0.6.2Safety mechanism for write operations
Connector Export/Import/Forkv0.7Distribution mechanism
OAuth 2.0v0.7Protocol implementation
MCP Server Exportv0.7Interoperability (depends on MCP adoption)
Database Connectorv0.8Direct DB access, connection pools
Message Pushv0.8Notification channels
RBACv0.8Access control, governance
Operation Audit Logv0.8Compliance
Sandbox Hardeningv0.9Security isolation
Observability (OTel, circuit breaker)v0.9Production operations
Connector Analyticsv0.9Usage tracking
Docker Composev0.9Deployment
Admin Dashboardv1.0Management UI
Scheduled Jobs / Webhooksv1.0Automation triggers
Batch Executionv1.0Enterprise-scale processing
Embeddable Widget / iframev1.0Delivery mode
Enterprise Securityv1.0Compliance (encryption, IP whitelisting)

Tailwind

FeatureVersionRelationship
AI Connector Builderv0.6.3Smarter models → better builder output
AI Connector Generation (OpenAPI)v1.0Same — models understand API specs better → more accurate auto-generation

Frozen (shipped, maintain only)

FeatureVersionStatus
ReAct Agentv0.1Shipped, working
DAG Planning / Re-Planningv0.1, v0.5Shipped, working
Memory (Window, Summary, Compact)v0.2, v0.5Shipped, working
RAG pipeline (embedding, vector store, chunking, hybrid retrieval)v0.5Shipped, working
Grounded Generationv0.5Shipped, working
ContextGuard / Pinned Messagesv0.5Shipped, working

Consider (deferred indefinitely)

FeatureOriginal versionReason deferred
Multi-Agent Orchestrationv1.0Providers building natively
Semantic Memory StoreBacklogContext windows growing; providers adding native memory
Memory LifecycleBacklogSame as above

Implications

  1. Don’t go back to v0.5 features. Bug fixes yes, new capabilities no.
  2. Connector platform is the core investment. v0.6–v0.8 should receive the majority of engineering time.
  3. Enterprise engineering (RBAC, audit, security, deployment) is the moat. These are boring but defensible.
  4. Re-evaluate annually. If model progress stalls or a “frozen” feature turns out to still have significant gaps, reconsider.